The word “agent” is now doing the work of about six different ideas at once. A chatbot on a pricing page is an “agent.” A script that files an expense report is an “agent.” The thing a vendor is trying to sell you for the price of a full-time hire is an “agent.” So is the open-source project a sixteen-year-old shipped over a weekend.
When one word covers that much ground, it stops carrying information. And if you are an operator deciding where to spend real money and real risk, a fuzzy word is expensive. You end up buying a demo and inheriting a project.
So let me give you a definition you can actually use, and a way to picture it that holds up under pressure. The picture is a musician. Once you see an agent as a performer instead of a piece of software, almost every hard question about deploying one answers itself.
A model is a musician with no instrument
A model alone is a brilliant conversationalist. That is a chatbot. Useful, sometimes magical, but it sits in the room and talks. To get an agent, you have to put an instrument in its hands.
The model is the musician. Whether it is Claude, GPT-5.5, Gemini, or GLM, it is raw musicianship: an ear, training, and the memory of, in some sense, every song ever written. Drop a chord in front of it and it knows what comes next.
But a musician sitting in an empty room with no instrument does not produce music. They produce potential. They can hum, they can describe what they would play, they can tell you what a great solo would sound like. They cannot actually play it, because playing requires touching something in the world.
This is the first thing operators get wrong. They evaluate the musician, asking “is the model smart enough?”, when the question that decides whether they get music is what instrument you handed it and what song it is playing. The model is the least interesting variable. It is also the one everyone argues about.
Tools are the instruments
An instrument is how a musician reaches the world. A guitar turns intention into sound. For a model, the instrument is a tool: a connection to something real it can actually operate. Send the email. Query the database. Pull the invoice. Update the CRM record. Search the case law. Move the file.
A model with no tools can tell you, in beautiful prose, exactly how it would reconcile your accounts. A model with a tool reconciles your accounts. The distance between describing the work and doing the work is the entire difference between a clever assistant and an agent.
Instruments have to be tuned, and they have to be the right ones. You do not hand a violinist a trumpet and call it a band. Each tool you give an agent is a specific capability you have decided it should have, wired into a specific system, with specific limits. The agent is only as good as its tools. A genius musician with a broken guitar makes worse music than a decent one with a great instrument. Most “AI projects” that die do not die because the model was dumb. They die because nobody built the tools, or the tools did not fit the systems the company actually runs on.
This is why I tell operators the boring parts are the hard parts. Choosing the model is an afternoon. The work is building reliable tools into your billing system, your document store, your matter-management software, your decade of accumulated exceptions. That is the instrument-making, and it is where the value is.
Context is knowing the song, and the room
Give a world-class jazz pianist the finest grand piano on earth and tell them nothing else. What do they play? They do not know. They do not know the song, the key, the tempo, who else is on stage, or whether this is a wedding or a wake.
Context is everything the agent knows going into the moment: the request in front of it, the relevant history, your company’s policies, the specific account it is working on, the last three things that happened. A musician reads the room and reads the chart. An agent reads its context.
Two things matter here that operators routinely confuse.
First, context is not memory. A musician can hold a whole evening’s set in their head; the agent gets handed a stack of sheet music for this song and then it is gone. If you want it to remember last week’s performance, you have to write that down and hand it back. Context is what is on the stand right now, not what the performer carries forever. Treating the context window like a brain is the single most common reason agents feel “forgetful” or inconsistent. They were never given the chart.
Second, more context is not automatically better. Bury the melody under two hundred pages of unrelated charts and even a great player loses the thread. The skill in deploying an agent is handing it exactly the right music: enough to play the part, not so much that the part gets lost. Curating that is a craft. It is also, again, the unglamorous work nobody demos.
The loop is what makes it a performer, not a piano player
One thing actually separates an agent from everything else.
A chatbot is a player piano. You feed it a roll, it plays the notes, it stops. One input, one output. Ask, answer, done. It does not listen to itself.
An agent is a live performer. It plays a phrase, hears how it landed, and decides what to play next. It takes an action with one of its tools, sees the result, and chooses the next move based on what actually happened. Pull the invoice, notice a line item is missing, go find it, then reconcile. The agent runs a loop: act, observe, decide, act again, until the song is finished.
That loop is the whole ballgame. It is why an agent can handle a task that does not go in a straight line, where step three depends on what step two turned up. A real performance is never identical twice, because the performer is responding to the room in real time. That responsiveness is the capability you are buying. It is also exactly what makes agents harder to trust than scripts, because a performer who improvises can improvise wrong.
Which brings us to the parts everyone wants to skip and absolutely cannot.
Permissions are the venue rules
You would not hand a session musician the keys to the building, the till, and the master tapes on day one. You decide what they are allowed to touch.
Permissions are the same decision for an agent. What systems can it reach? Can it read the data or also change it? Can it send the email to a customer, or only draft it for someone to send? Can it move ten dollars or ten million? An agent without scoped permissions is a stranger you have handed the keys to. An agent with the right permissions is a trusted performer who plays their part and does not wander into the vault.
Operators who get burned almost always got the permissions wrong: too broad, too fast, with no one watching the first few sets. “Trust the AI more” is the wrong fix. Treat it like any new hire, with a tight scope that earns expansion by performing.
Evals are the audition
No serious venue puts a musician on stage without hearing them play first. You audition. You hand them the actual material and you listen. Can they hold the tempo, hit the changes, recover when something goes sideways?
Evals are the audition for an agent, and the industry skips them constantly. People watch one impressive demo, one great solo at the audition, and book the act for a residency. Then they are shocked when the agent flubs the fortieth performance on a Tuesday with a weird input nobody rehearsed.
A demo proves the agent can do the task once. An eval proves it does the task reliably across the messy reality of your actual work: the malformed inputs, the exception that shows up in one matter out of fifty. This is the entire reason we built BORE, our production eval. The question that matters is not “is the model impressive?” It is “can it do your job, on your inputs, at the rate you need?” You do not learn that from a sales call. You learn it from an audition built out of your own material. Anyone deploying agents without one is gambling, and calling it strategy.
Human review is the producer in the booth
Even the best performers play to a producer. Someone in the booth is listening, ready to stop the take, flag the bad note, decide what ships and what gets cut. The performer plays; the producer is accountable for what leaves the studio.
This is the part I push hardest on, because it is usually framed as a weakness. “Well, a human still has to check it.” That framing is backwards. Human-in-the-loop is a design decision, and it is the most important one in the whole build. The question is never whether a human is involved; it is where, and on what. The agent drafts the contract; the lawyer approves the language that goes out. The agent reconciles the ledger; the controller signs off on the entries above a threshold. The agent routes the case; a person owns the call when it is genuinely ambiguous.
Designing those review gates well, knowing which notes need the producer and which the performer can ship alone, is most of what separates an agent that creates leverage from one that creates liability. Get it right and your best people stop doing the rote work and start doing only the judgment. Get it wrong and you have either automated your mistakes or recreated the bottleneck you were trying to remove.
Why this is the whole point: the model is not the system
Put it together and you can see why most AI projects fail in a way most AI demos never reveal.
A demo shows you a brilliant musician playing a flawless solo in a quiet studio. A deployment is a working performer who has the right instruments, knows the song, plays live to a room that keeps changing, stays inside the venue rules, passed a real audition, and answers to a producer who owns what ships. The first is a model. The second is a system. The model is maybe ten percent of it. The other ninety percent is the tools, the context, the permissions, the evals, and the review gates: the unglamorous engineering that nobody puts in a keynote and everybody needs in production.
When someone sells you “an AI agent,” ask which one they mean: the solo in the quiet room, or the performer who survives the residency. The price tags look similar. The outcomes do not.
And now the band
Everything so far is one musician. The frontier, and increasingly the reality inside the companies we build for, is the band.
A single performer can carry a song. They cannot be a whole orchestra. So you do what music has always done: you assign parts. The bassist holds the bottom. The drummer keeps time. The horns take the melody. Each player is a specialist agent with its own tools, its own sheet music, its own narrow excellence. None of them is trying to play everything, which is exactly why the whole thing sounds good.
A real business process is an arrangement, not a solo. An invoice comes in. One agent reads and extracts it. Another checks it against the contract. Another flags the exception for a human. Another posts the approved entry and updates the customer. No single agent is brilliant at all four; each is excellent at one, and they hand off cleanly, in time, like sections of a band trading phrases.
And a band needs a conductor. Somebody has to count it in, hold the tempo, bring sections in and out, and stop the whole thing when it falls apart. In a multi-agent system that is the orchestration layer: the part that decides which agent plays when, passes the work between them, and keeps the arrangement coherent. It is the least flashy seat in the house and the one that turns a room full of talented players into music instead of noise.
That conductor’s stand is the seat the operator should care about most. The musicians, the models, are a commodity; you can hire brilliant ones from anyone. The instruments can be bought. The arrangement is the thing that is yours: knowing your business well enough to score it, deciding which parts get automated and which stay human, building the handoffs so a hundred performances a day stay in tune. That arrangement, written down in software around your actual workflow, is what we call a custom AI operating system. It is the asset. It is the moat that is still working while you sleep.
The model is the musician. The tools are the instruments. The context is the song. The permissions are the venue. The evals are the audition. The review is the producer. And the system you actually want is a band: one that knows your music, plays it the same way every night, and never misses the gig.
Most companies are still arguing about which musician to hire. The ones pulling ahead are writing the arrangement.