Immediate Constancy: Measuring How A lot of Your Intent an AI Agent Really Executes

Spotify simply shipped “Prompted Playlists” in beta. I constructed a number of playlists and found that the LLM behind the agent tries to meet your request, however fails as a result of it doesn’t know sufficient however gained’t admit it. Right here’s what I imply: certainly one of my first playlist prompts was “songs in a minor key inside rock”. The playlist was swiftly created. I then added the caveat “and no tune ought to have greater than 10 million performs”. The AI agent bubbled up an error explaining that it didn’t have entry to whole play counts. It additionally surprisingly defined that it didn’t have entry to some different issues like musical keys, regardless that it had claimed to make use of that within the playlist’s development. The agent was utilizing its LLM’s information of what key a sure tune was in and including songs accordingly to its reminiscence. A detailed inspection of the playlist confirmed a number of songs that weren’t in a minor key in any respect. The LLM had, in fact, hallucinated this data and proudly displayed it as a legitimate match to a playlist’s immediate.

Clearly, a playlist creator is a reasonably low-stakes AI agent functionality. The playlist it made was nice! The difficulty is it solely actually used about 25% of my constraints as validated enter. The remaining 75% of my constraints have been simply guessed by the LLM and the system by no means informed me till I dug in deeper. This isn’t a Spotify drawback; it’s an every-agent drawback. 

Three Propositions

To exhibit this idea of immediate constancy extra broadly, I have to make these three propositions:

1. Any AI agent’s verified knowledge layer has a restricted or finite capability. An agent can solely question the instruments it’s been given, and people instruments expose a hard and fast set of fields with finite decision. You’ll be able to enumerate each area within the schema and measure how a lot each narrows the search. A reputation rating eliminates some fraction of candidates. A launch date eliminates one other. A style tag eliminates extra. Add up how a lot narrowing all of the fields can do collectively and also you get a tough quantity: the utmost quantity of filtering the agent can show it did. I’ll name that quantity I_max.
2. Consumer intent expressed in pure language is successfully unbounded. An individual can write a immediate of arbitrary specificity. “Create a playlist with songs which are bass-led in minor key, post-punk from Manchester, recorded in studios with analog tools between 1979 and 1983 that influenced the gothic rock motion however by no means charted.” Each clause narrows the search. Each adjective provides precision. There is no such thing as a ceiling on how particular a person’s request could be, as a result of pure language wasn’t designed round database schemas.
3. Following immediately from the primary two: for any AI agent, there exists some extent the place the person’s immediate asks for greater than the info layer can confirm. As soon as a immediate calls for extra narrowing than the verified fields can present, the remaining work has to come back from someplace. That someplace is the LLM’s normal information, sample matching, and inference. The agent will nonetheless ship a assured outcome. It simply can’t show all of it. Not as a result of the mannequin is poorly constructed, however as a result of the maths doesn’t enable the rest.

This isn’t a high quality drawback, however a structural one. A greater mannequin doesn’t increase the ceiling. Higher fashions do get higher at inferring and filling in the remainder of the person’s wants. Nonetheless, solely including extra verified knowledge fields raises this ceiling, and even then, every new area gives diminishing returns as a result of fields are correlated (style and power aren’t impartial, launch date and tempo traits aren’t impartial). The hole between what language can specific and what knowledge can confirm is everlasting.

The Downside: Brokers Don’t Report Their Compression Ratio

Each AI agent with entry to instruments and expertise does the identical factor: it takes your request, decomposes that request right into a set of actions, executes these actions, infers concerning the output of these actions, after which presents a unified response. 

This decomposition from request to motion truly erodes the that means between what it’s you’re asking for and what the AI agent responds with. The narration layer of the AI agent flattens what it’s you requested and what was inferred right into a single response. 

The issue is that as a person of an AI agent, you haven’t any strategy to know what fraction of your enter was used to set off an motion, what fraction of the response was grounded in actual knowledge, and what fraction was inferred from the actions that the agent took. It is a drawback for playlists as a result of there have been songs that have been in a serious key, after I had explicitly requested it to solely comprise songs in a minor key. That is much more of an issue when your AI agent is classifying monetary receipts and transactions. 

We’d like a metric for measuring this. I’m calling it Immediate Constancy. 

The Metric: Immediate Constancy

Immediate Constancy for AI brokers is outlined by the constraints you give to the agent when asking it to carry out some motion. Every constraint inside a immediate narrows the potential paths that the agent can take by some measurable quantity. A naïve method to calculating constancy can be to depend every constraint, add up those which are verifiable, and those which are inferred. The issue with that method is that every constraint is weighted the identical. Nonetheless, knowledge is commonly skewed closely inside actual life datasets. A constraint that eliminates 95% of the catalog is doing vastly extra work than one which eliminates 20%. Counting every constraint the identical is unsuitable.

Due to this fact, we have to correctly weight every constraint in line with the work it does filtering the dataset. Logarithms obtain that weighting. The bits of data in a immediate could be outlined as “-log2(p)” bits the place p is the surviving fraction of data from the constraints or fillers you’ve utilized. 

In every agent motion, every constraint can solely be a) verified by device calls or b) inferred by the LLM. Immediate constancy measures the ratio of constraints between these two choices. 

Immediate Constancy has a variety of 0 to 1. An ideal 1.0 implies that each a part of your request was backed by actual knowledge. A constancy of 0.0 implies that the whole output of the AI agent was pushed by its inside reasoning or vibes. 

Spotify’s system above all the time studies an ideal 1.0 on this scenario. In actuality, the immediate constancy of the playlist creation was round 25% – two constraints (underneath 4 minutes and recorded earlier than 2005) have been fulfilled by the agent, the remainder have been inferred from the agent’s current (and doubtlessly defective) information and recall. At scale and utilized to extra impactful issues, falsely reporting a excessive immediate constancy turns into an enormous drawback.

What Constancy Really Means (and Doesn’t Imply)

In audio methods, “constancy” is a measure of how faithfully the system reproduces the unique sign. Excessive constancy doesn’t assure that the music itself is sweet. Excessive constancy solely ensures that the music sounds the way it did when it was recorded. Immediate constancy is identical concept: how a lot of your unique intent (sign) was faithfully fulfilled by the agentic system.

Excessive immediate constancy implies that the system did what you requested and you’ll PROVE it. A low immediate constancy means the system most likely did one thing shut to what you wished, however you’ll should evaluation it (listening to the entire playlist) to make sure that it’s true. 

Immediate Constancy is NOT an accuracy rating. It can’t inform you that “75% of the songs in a playlist match your immediate”. A playlist with a 0.25 constancy might be 100% excellent. The LLM may need nailed each single inference about every tune it added. Or, half the songs might be unsuitable. You don’t know. You’ll be able to’t know till you take heed to all of the songs. That’s the purpose of a measurable immediate constancy. 

As an alternative immediate constancy measures how a lot of the outcome you’ll be able to TRUST WITHOUT CHECKING. In a monetary audit, if 25% of the road gadgets have receipts and 75% of the road gadgets are estimates, the entire invoice may nonetheless be 100% correct, however your CONFIDENCE in that whole is essentially completely different than an audit with each single line merchandise supported by a receipt. The excellence issues as a result of there are domains the place ‘simply belief the vibes’ is ok (music) and domains the place it isn’t (medical recommendation, monetary steerage, authorized compliance).

Immediate constancy is extra like a measurement of the documentation price given quite a lot of constraints, not the error price of the response itself. 

Virtually in our Spotify instance: as you add extra constraints to your playlist immediate, the immediate constancy drops, the playlist turns into much less of a exact report and extra of a advice. That’s completely superb, however the person ought to be knowledgeable about which they’re getting. Is that this playlist precisely what I requested for? Or did you make one thing work to meet the aim that I gave you? Surfacing that metric to the person is crucial for constructing belief in these agentic methods.

The Case Research: Reverse-Engineering Spotify’s AI Playlist Agent

Spotify’s Prompted Playlists function is what began this exploration into immediate constancy. Let’s dive deeper into how these work and what I did to discover this functionality simply from the usual immediate enter area.

Prompted Playlists allow you to describe what you need in pure language. For instance, on this playlist, the immediate is solely “rock songs in minor keys, underneath 4 minutes, recorded earlier than 2005, that includes bass strains as a lead melodic aspect”. 

Usually, to make a playlist, you’d must comb by way of hours of music to land on precisely what you wished to make. This playlist is 52 minutes lengthy and took solely a minute to generate. The attraction right here is apparent and I actually take pleasure in this function. With out having to know all the important thing rock artists, I could be launched to the music and discover it extra rapidly and extra simply. 

Sadly, the official documentation from Spotify may be very gentle. There are virtually no particulars about what the system can or can’t do, what metadata it keys off of, neither is there any knowledge mapping obtainable. 

Utilizing a easy method, nevertheless, I used to be capable of map what I imagine is the total knowledge contract obtainable to the agent over the course of 1 night (all from my sofa watching the Sopranos, naturally).

The Method: Inconceivable Constraints as a Forcing Operate

Because of how Spotify architected this playlist-building agent, when the agent can’t fulfill a request, the error messages could be influenced to disclose architectural particulars which are in any other case not obtainable. Whenever you discover a constraint that the agent can’t construct off of, it can error and you’ll leverage that to know what it CAN do. I’ll use this because the fixed to probe the system. 

In our instance playlist above, Minor Keys & Bass Strains, including the unlock phrase “with lower than 10 million streams” acts as a circuit breaker for the agent, signalling that it can’t fulfill the customers’ request. With this phrase, you’ll be able to discover the chances by altering different facets of the immediate again and again till you’ll be able to see what the agent has entry to. Amassing the responses, asking overlapping questions, and reviewing the responses means that you can construct a foundational understanding of what’s obtainable for the agent. 

What I Discovered: The Three-Tier Structure

Spotify Prompted Playlist agent has a wealth of information obtainable to it. I’ve separated it into three tiers: musical metadata, user-based knowledge, and LLM inference. Past that, it seems that Spotify has excluded numerous knowledge sources from its agent both as a product selection or as a “get this out the door” selection. 

• Tier 1
  • Verified monitor metadata: length, launch date, reputation, tempo, power, specific, style, language
• Tier 2
  • Verified person behavioral knowledge: play counts, skip counts, timestamps, recency flags, ms performed, supply, interval analytics (40+ fields whole)
• Tier 3
  • LLM inference: key/mode, danceability, valence, acousticness, temper, instrumentation — all inferred from normal information, narrated as if verified
• Deliberate exclusion:
  • Spotify’s public API has audio options (danceability, valence, and so forth.) however the agent doesn’t have entry. Maybe a product selection, not technical limitation.

A full checklist of obtainable fields is included on the backside of this submit. 

The Behavioral Findings

The agent demonstrated surprisingly resilient conduct to ambiguous requests and conflicting directions. It generally reported that it was doublechecking numerous constraints and fulfilling the customers’ request. Nonetheless, whether or not these constraints have been truly checked towards a validated dataset or not was not uncovered. 

When the playlist agent can get an in depth, however not actual, match to the constraints listed within the immediate, it runs a “associated” question and silently substitutes the outcomes from that question as legitimate outcomes for the unique request. This dilutes the belief within the system since a immediate requesting ONLY bass-driven rock music in a playlist may collect non-bass-driven rock music in a playlist, doubtless dissatisfying the person.

There does seem like a “certainty threshold” that the agent just isn’t snug crossing. For instance, this whole exploration was primarily based on the “lower than 10 million performs” unlock phrase. When this occurs, the agent would expose only a handful of fields it had entry to each time. This checklist of fields would change from immediate to immediate, even when the immediate was the identical between runs of the immediate. That is basic LLM non-determinism. With a view to enhance belief within the system, exposing what the agent DOES have entry to in a simple approach tells the human precisely what they will and can’t ask about. 

Lastly, when these two kinds of knowledge are combined, the agent just isn’t clear about which songs it has used verified knowledge for and which it has used inferred knowledge for. Each verified and inferred selections are combined and offered with an identical authority within the music notes. For instance, in case you craft a prompted playlist about your individual person data (“songs I’ve skipped greater than 30 occasions with a punchy bass-driven melody”), the agent will add actual knowledge (“you skipped this tune 83 occasions final yr!”) proper subsequent to inferred information (“John Deacon’s bass line instructions consideration all through this tune”). To be clear, I’ve not skipped any Queen songs 83 occasions to my information. However the AI agent doesn’t have a “bass_player” area anyplace in its obtainable knowledge to question towards. The AI is aware of that Queen generally has a robust bass line of their songs and the information of John Deacon as Queen’s bass guitarist permits its LLM to deduce that it’s his bass line that brought about the tune to be added to the playlist.

Making use of the Math: Two Playlists, Two Constancy Scores

Let’s apply this immediate constancy idea to instance playlists. I don’t have full entry to the Spotify music catalog so I’ll be utilizing instance survivorship numbers from our standards filters in our constancy bit computations. The system is identical at each step: bits = −log₂(p) the place p is the estimated fraction of the catalog that survives the filter being utilized.

“Minor Bass Melodies” — The Assured Phantasm

This playlist is the one with Queen. “A playlist of rock music, all in minor key, underneath 4 minutes of playtime, launched pre-2005, and bass-led”. I’ll apply our system and use the bits of data I’ve from every step to assist compute the immediate constancy.

Length < 4 minutes

• Estimate: ~80% of tracks are underneath 4 minutes → p = 0.80

• This barely narrows something, which is why it contributes so little

Launch date earlier than 2005

• Estimate: ~30% of Spotify’s catalog is pre-2005 (the catalog skews closely towards latest releases) → p = 0.30

• Extra selective — eliminates 70% of the catalog

Minor key

• Estimate: ~40% of widespread music is in a minor key → p = 0.40

• Average selectivity, however that is totally inferred — the agent confirmed key/mode just isn’t a verified area

Bass-led melodic aspect

• Estimate: ~5% of tracks function bass because the lead melodic aspect → p = 0.05

• By far essentially the most selective constraint. This single filter does extra work than the opposite three mixed. And it’s 100% inferred.

Totals:

These survival fractions are estimates. Nonetheless, the structural level holds no matter actual numbers: essentially the most selective constraint is the least verifiable, and that’s not a coincidence. The issues that make a immediate attention-grabbing are virtually all the time the issues an agent has to guess at.

“Skipped Songs” — The Sincere Playlist

This immediate may be very straight ahead: “A playlist of songs I’ve skipped greater than 5 occasions”. That is very simple to confirm and the agent will lean into the info it has entry to.

Skip depend > 5

• Estimate: ~10% of tracks in your library have been skipped greater than 5 occasions → p = 0.10

• That is the one constraint, and it’s a verified area (user_skip_count)

Totals:

The Structural Perception

The attention-grabbing half about immediate constancy is obvious in every playlist: the “most attention-grabbing” immediate is the least verifiable. A playlist with all my skipped songs is trivially simple to implement however Spotify doesn’t wish to present it. In any case, these are all songs I typically don’t favor to take heed to, therefore the skips. Equally, publish date being earlier than 2005 may be very simple to confirm, however the resultant playlist is unlikely to be attention-grabbing to the typical person.

The bass-line constraint although may be very attention-grabbing for a person. Constraints like these are the place the Prompted Playlist idea will shine. Already right this moment I’ve created and listened to 2 such playlists generated from only a idea of a tune that I wished to listen to extra of. 

Nonetheless, the idea of a “bass-driven” tune is tough to quantify, particularly at Spotify’s scale. Even when they did quantify it, I’d ask for “clarinet jazz” the subsequent day and so they’d all should get again to work discovering and labeling these songs. And that is in fact the magic of the Prompted Playlist function.

Validation: A Managed Agent

The Spotify examples are compelling, however I don’t have direct entry to the schema, the instruments, and the agentic harness itself. So I constructed a film advice agent to be able to take a look at this principle inside a extra managed setting.

https://github.com/Barneyjm/prompt-fidelity 

The film advice agent is constructed with the TMDB API that gives the verified layer. Fields within the schema are style, yr, score, runtime, language, solid, and director. All the opposite constraints like temper, tone, and pacing usually are not verified knowledge and are as an alternative sourced from the LLM’s personal information of films. Because the agent fulfills a person’s request, the agent information its knowledge sources as both verified or inferred and scores its personal response. 

The Boring Immediate (F = 1.0)

We’ll begin with a “boring” immediate: “Motion motion pictures from the Eighties rated above 7.0”. This gives the agent three constraints to work with: style, date vary, and score. All these constraints correspond to verified knowledge values throughout the database. 

If I run this by way of the take a look at agent, I see the excessive constancy pops out naturally as a result of every constraint is tied to verified knowledge. 

Each outcome right here is verifiably appropriate. The LLM made zero judgement calls as a result of it had knowledge it might base its response on for every constraint.

The Vibes Immediate (F = 0.0)

On this case, I’ll search for “motion pictures that really feel like a wet Sunday afternoon”. No constraints on this immediate align to any verified knowledge in our dataset. The work required of the agent falls totally on its LLM reasoning off its current information of films.

The suggestions are defensible and are actually good motion pictures however they don’t seem to be verifiable in line with the info we’ve got entry to. With no verified constraints to anchor the search, the candidate pool was the whole TMDb catalog, and the LLM needed to do all of the work. Some picks are nice; others are the mannequin reaching for obscure movies it isn’t assured about.

The Takeaway

This take a look at film advice agent verifies the immediate constancy framework as a robust strategy to expose how an agent’s interpretation of a customers’ intent pushes its response right into a precision device or a advice engine. The place the response lands between these two choices is vital for informing customers and constructing belief in agentic methods. 

The Constancy Frontier

To make this concrete: Spotify’s catalog accommodates roughly 100 million tracks. How a lot whole data your immediate wants to hold to slender the catalog right down to your playlist I’ll name I_required.

To pick out a 20-song playlist from that catalog, you want roughly 22 bits of selectivity (log₂ of 100 million divided by 20).

The verified fields (length, launch date, reputation, tempo, power, style, specific flag, language, and the total suite of person behavioral knowledge) have a mixed capability that tops out at roughly 10 to 12 bits, relying on the way you estimate the selectivity of every area. After that, the verified layer is exhausted. Each further little bit of specificity your immediate calls for has to come back from LLM inference. I’ll name this most, I_max

That provides you a constancy ceiling for any immediate:

And the constancy ceiling for any playlist:

For the Spotify agent, a maximally particular immediate that absolutely defines a playlist can’t exceed roughly 55% constancy. The opposite 45% is structurally assured to be inference. For easier prompts that don’t push previous the verified layer’s capability, constancy can attain 1.0. However as prompts get extra particular, constancy drops, not progressively however by necessity.

This defines what I’m calling the constancy frontier: the curve of most achievable constancy as a operate of immediate specificity. Each agent has one. It’s computable prematurely from the device schema. Easy prompts sit on the left of the curve the place constancy is excessive. Artistic, particular, attention-grabbing prompts sit on the fitting the place constancy is structurally bounded beneath 1.0.

The uncomfortable implication is that the prompts customers care about most (those that really feel private, particular, and tailor-made) are precisely those that push previous the verified layer’s capability. Probably the most attention-grabbing outputs come from the least devoted execution. And essentially the most boring prompts are essentially the most reliable. That tradeoff is baked into the maths. It doesn’t go away with scale, higher fashions, or greater databases. It solely shifts.

For anybody constructing brokers, the sensible takeaway is that this: you’ll be able to compute your individual I_max by auditing your device schema. You’ll be able to estimate the standard specificity of your customers’ prompts. The ratio tells you the way a lot of your agent’s output is structurally assured to be inference. That’s a quantity you’ll be able to put in entrance of a product crew or a threat committee. And for brokers dealing with coverage questions, medical data, or monetary recommendation, it means there’s a provable decrease certain on how a lot of any response can’t be grounded in retrieved knowledge. You’ll be able to shrink it. You can’t get rid of it.

The Broader Software: Each Agent Has This Downside

This isn’t a Spotify drawback. It is a drawback for any system the place an LLM orchestrates device calls to reply a person’s query.

Take into account Retrieval Augmented Era (RAG) methods, which energy most enterprise AI knowledge-base deployments right this moment. When an worker asks an inside assistant a coverage query, a part of the reply comes from retrieved paperwork and half comes from the LLM synthesizing throughout them, filling gaps, and smoothing the language into one thing readable. The retrieval is verified. The synthesis is inferred. And the response reads as one seamless paragraph with no indication of the place the seams are. A compliance officer studying that reply has no strategy to know which sentence got here from the enterprise coverage doc and which sentence the mannequin invented to attach two paragraphs that didn’t fairly match collectively. The constancy query is an identical to the playlist query, simply with larger stakes.

Coding brokers face the identical decomposition. When an AI generates a operate, a few of it might reference established patterns from its coaching knowledge or documentation lookups, and a few of it’s novel era. As extra manufacturing code is written by AI, surfacing that ratio turns into an actual engineering concern. A operate that’s 90% grounded in well-tested patterns carries completely different dangers than one which’s 90% novel era, even when each go the identical take a look at suite right this moment.

Customer support bots stands out as the highest-stakes instance. When a bot tells a buyer what their refund coverage is, that reply ought to be drawn immediately from coverage paperwork, full cease. Any inferred or synthesized content material in that response is a legal responsibility. The silent substitution conduct noticed in Spotify (the place the agent ran a close-by question and narrated it as if it fulfilled the unique request) can be genuinely harmful in a customer support context. Think about a bot confidently stating a return window or protection time period that it inferred somewhat than retrieved.

The overall type of immediate constancy applies to all of those:

Constancy = bits of response grounded in device calls / whole bits of response

The onerous half, and more and more the core problem of AI engineering work, is defining what “bits” means in every context. For a playlist with discrete constraints, it’s clear. Totally free-text era, you’d must decompose a response into particular person claims and assess each, which is nearer to what factuality benchmarks already attempt to do, simply reframed as an information-theoretic measure. That’s a tough measurement drawback, and I don’t declare to have solved it right here.

However I feel the framework has worth even when actual measurement is impractical. If the individuals constructing these methods are fascinated with constancy as a design constraint (what fraction of this response can I floor in device calls, and the way do I talk that to the person?) the outputs will likely be extra reliable whether or not or not anybody computes a exact rating. The aim isn’t a quantity on a dashboard. The aim is a psychological mannequin that shapes how we construct. 

The Complexity Ceiling

Each agent has a complexity ceiling. Easy lookups (what’s the play depend for this monitor?) are primarily free. Filtering the catalog towards a set of field-level predicates (present me every thing underneath 4 minutes, pre-2005, reputation beneath 40) scales linearly and runs quick. However the second a immediate requires cross-referencing entities towards one another (does this monitor seem in additional than three of my playlists? was there a year-long hole someplace in my listening historical past?) the associated fee jumps quadratically, and the agent both refuses outright or silently approximates.

That silent approximation is the attention-grabbing failure mode. The agent follows a sort of precept of least computational motion: when the precise question is simply too costly, it relaxes your constraints till it finds a model it may well afford to run. You requested for a selected valley within the search area; it rolled downhill to the closest one as an alternative. The result’s an area minimal, shut sufficient to look proper, low-cost sufficient to serve, however it’s not what you requested for, and it doesn’t inform you the distinction.

This ceiling isn’t distinctive to Spotify. Any agent constructed on listed database lookups will hit the identical wall. The boundary sits proper the place queries cease being decomposable into impartial WHERE clauses and begin requiring joins, full scans, or aggregations throughout your total historical past. Under that line, the agent is a precision device. Above it, it’s a advice engine carrying a precision device’s garments. The query for anybody constructing these methods isn’t whether or not the ceiling exists (it all the time does) however whether or not your customers know the place it’s.

What to Do About It: Design Suggestions

If immediate constancy is an actual and measurable property of agentic methods, the pure query is what to do about it. Listed here are 5 suggestions for anybody constructing or deploying AI brokers with device entry.

• Report constancy, even roughly. Spotify already exhibits audio high quality as a easy indicator (low, regular, excessive, very excessive) if you’re streaming music. The identical sample works for immediate constancy. You don’t want to indicate the person a decimal rating. A easy label (“this playlist carefully matches your immediate” versus “this playlist is impressed by your immediate”) can be sufficient to set expectations appropriately. The distinction between a precision device and a advice engine is ok, so long as the person is aware of which one they’re holding.
• Distinguish grounded claims from inferred ones within the UX. This may be refined. A small icon, a slight colour shift, a footnote. When Spotify’s playlist notes say “86 skips” that’s a truth from a database. Once they say “John Deacon’s bass line drives the entire monitor” that’s the LLM’s normal information. Each are offered identically right this moment. Even a minimal visible distinction would let customers calibrate their belief per declare somewhat than trusting or distrusting the whole output as a block.
• Disclose substitutions explicitly. When an agent can’t fulfill a request precisely however can get shut, it ought to say so. “I couldn’t filter on obtain standing, so I discovered songs from albums you’ve saved however haven’t appreciated” preserves belief way over silently serving a close-by outcome and narrating it as if the unique request was fulfilled. Customers are forgiving of limitations. They’re much much less forgiving of being misled.
• Present deterministic functionality discovery. Once I requested the Spotify agent to checklist each area it might filter on, it produced a special reply every time relying on the context of the immediate. The LLM was reconstructing the sphere checklist from reminiscence somewhat than studying from a hard and fast reference. Any agent that exposes filtering or querying capabilities to customers ought to have a steady, deterministic strategy to uncover these capabilities. A “present me what you are able to do” command that returns the identical reply each time is desk stakes for person belief.
• Audit your individual agent with this system earlier than your customers do. The methodology on this piece (pairing inconceivable constraints with goal fields to pressure informative refusals) is a general-purpose audit method that works on any agent with device entry. It took one night and a couple of dozen prompts to map Spotify’s full knowledge contract. Your customers will do the identical factor, whether or not you invite them to or not. The query is whether or not you perceive your individual system’s boundaries earlier than they do.

Closing

Each AI agent has a constancy rating. Most are decrease than you’d anticipate. None of them report it.

The methodology right here (utilizing inconceivable constraints to pressure informative refusals) isn’t particular to music or playlists. It really works on any agent that calls instruments. If the system can refuse, it may well leak. If it may well leak, you’ll be able to map it. A dozen well-crafted prompts and a night of curiosity is all it takes to know what a manufacturing agent can truly do versus what it claims to do.

The maths generalizes too. Weighting constraints by their selectivity somewhat than simply counting them reveals one thing {that a} naïve audit misses: the constraints that make a immediate really feel private and particular are virtually all the time those the system can’t confirm. Probably the most attention-grabbing outputs come from the least devoted execution. That stress doesn’t go away with higher fashions or greater databases. It’s structural.

As AI brokers develop into the first approach individuals work together with knowledge methods (their music libraries right this moment, their monetary accounts and medical information tomorrow) customers will probe boundaries. They’ll discover the gaps between what was promised and what was delivered. They’ll uncover that the assured, well-narrated response was partially grounded and partially invented, with no strategy to inform which components have been which.

The query isn’t whether or not your agent’s constancy will likely be measured. It’s whether or not you measured it first.

Bonus: Prompts Value Attempting (If You Have Spotify Premium)

As soon as the schema, you’ll be able to write prompts that floor genuinely shocking issues about your listening historical past. These all labored for me with various levels of tweaking:

The Relationship Post-mortem

• “Songs the place my skip depend is larger than my play depend”
• Honest warning: this one might trigger existential discomfort (you skip these songs for a motive!)

Love at First Hear

• “Songs the place I saved them inside 24 hours of my first play, sorted by oldest first”
• A chronological timeline of tracks that grabbed you instantly

The Lifecycle

• “Songs I first ever performed, sorted by most performs”
• Your origin story on the platform

The Marathon

• “Songs the place my whole ms_played is highest, convert to hours”
• Not most performs — most whole time. A unique and infrequently shocking checklist

The Longest Relationship

• “Songs with the smallest hole between first play and most up-to-date play, with at the very least 50 performs, ordered by earliest first pay attention”

The One-Week Obsessions

• “Songs I performed greater than 10 occasions in a single week after which by no means touched once more”
• Your former obsessions, fossilized. This was like a time machine for me.

The Time Capsule

• “One tune from annually I’ve been on Spotify — the tune with essentially the most performs from that yr”

The Earlier than and After

• “Two units: my 10 most-played songs within the 6 months earlier than [milestone date] and my 10 most-played within the 6 months after”
• Plug in any date that mattered — a transfer, a brand new job, a breakup, and even Covid-19 lockdown

The Soundtrack to a Yr

• “Choose the yr the place my whole ms_played was highest. Construct a playlist of my high songs from that yr”

What Didn’t Work (and Why)

• Comeback Story (year-long hole detection): “Songs I rediscovered after a year-long hole in listening”
  • agent can’t scan full play historical past for gaps. Snapshot queries work, timeline scans don’t.
• Seasonal patterns (solely performed in December): “Songs I solely performed in December however by no means every other month”
  • proving common negation requires full scan. Identical basic limitation.
• Derived math (ms_played / play_count): “Songs the place my common pay attention time is underneath 30 seconds per play”
  • agent struggles with computed fields. Stick with uncooked comparisons.
• These failures map on to the complexity ceiling — they require O(n²) or full-scan operations the agent can’t or isn’t allowed to carry out.

Ideas

• Reference area names immediately when the agent misinterprets pure language
• Begin broad and tighten. Unfastened constraints succeed extra usually
• “In the event you can’t do X, inform me what you CAN do” is the common audit immediate

Monitor Metadata

Audio Options (Partial)

Consumer Behavioral Information

Supply & Context

Interval Analytics (Time-Windowed)

Question / Search Fields

Confirmed Unavailable