From the outside, an agent looks like it is improvising. It reads a message, decides what to do, picks a tool, checks the result and writes back. Every conversation seems different.
Read enough of its logs and another picture appears. Where is my order. Cancel this one. Change the address on that one. Can I return this. The wording changes every time; the job underneath doesn't.
The shape of agentic work
Most of what a production agent does falls into a few repeated jobs, with a long tail of everything else behind them. On τ-bench retail, 5 repeated jobs, found in raw agent logs with no task list, cover 85% of the agent's writes.
The repeated jobs are where the volume is. The long tail is where the judgment is. Today's agents treat both the same way: every message gets the whole loop of read, reason, call, check and reply.
Why the loop adds up
- Every turn is a model call, and every call reads the whole conversation again.
- Reasoning varies. The same job can go a little differently each time.
- Volume compounds. A job your agent does a thousand times a day is a thousand fresh decisions.
The wording changes every time. The job underneath doesn't.
Two kinds of work, two paths
The answer isn't a bigger model. It is giving the repeated jobs their own path, so your agent's model is spent where it matters. That's the split AgentCompile makes: known jobs run compiled, and anything new, unclear or unusual goes to your agent, unchanged, with the whole conversation so far.
A thought experiment
Pick any hour of your agent's logs and sort the conversations by what the agent actually did, ignoring the wording. A few piles grow tall fast: order status, refunds, address changes, cancellations. A scatter of one-offs sits beside them. That's the shape we keep seeing: a short head of repeated jobs, and a long tail of everything else.
Without AgentCompile
- The head and the tail get the same treatment.
- Volume in the head means volume in model calls.
- Variation in the head means variation in answers.
With AgentCompile
- The head runs compiled, the same way every time.
- Your agent's model is called for the tail.
- The tail is where your agent's judgment matters, and where it goes.
What this means for how you build
- Measure jobs, not messages. Two messages that look nothing alike are often the same job.
- Look at writes. What your agent changes tells you more about its jobs than what it says.
- Start from your own history. Your agent's jobs are its own; a pilot starts by showing you which repeat.
See what that did on τ-bench retail. Read the results