Quota discipline in an agentic loop

Agentic sessions are where allowances actually go. Not because the model is expensive per call, but because a loop that reads files, runs commands and reads the results accumulates an enormous bundle and re-sends it every turn.

Usefully, the habits that reduce consumption are the same ones that improve output. Both are the same problem: too much undifferentiated context.

Six habits, roughly in order of payoff

1. Scope the task before starting. "Fix the failing test in auth_test.go" costs a fraction of "the tests are failing, have a look". The second sends the agent exploring, and exploration is reading, and reading is the bill.

2. Let it plan first. A cheap planning pass that produces an approach you approve prevents an expensive execution pass in the wrong direction. Wrong direction is the most expensive thing in an agentic workflow because you pay for it twice — once to do it and once to undo it.

3. Start fresh more often than feels natural. Context does not improve with age; it accumulates. When a task is done, the next task rarely needs the previous one's forty tool results. A new session with a two-line summary is cheaper and usually sharper.

4. Put the stable facts in CLAUDE.md, not in every prompt. Build commands, conventions, where things live. Re-explaining your project each session is a recurring cost that a file removes once.

5. Match the model to the step. Bulk mechanical work — reading, summarising, classification, first drafts — does not need the most capable model. Save that for the reasoning. See Choosing a model.

6. Bound the loop. An agent asked to "keep going until it works" will. Say what done looks like and when to stop and ask.

Try this

Run one session with your usual habits and note where consumption lands. Then run a comparable task having changed only this much: state the scope precisely up front, and start from a fresh session. The difference is usually larger than any prompt tuning you could do.

What goes wrong

Letting the agent explore because it is easier than being specific. It is easier. It is also the single biggest cost multiplier available.

Never starting a new session. One conversation carried across an entire day is expensive at the end and worse at the end, both for the same reason.

Dumping files into context "so it has everything". Now every turn pays for everything.

Treating a limit as arbitrary. It is a fairly direct readout of how much you asked the model to read.

How to check it worked

After changing habits, compare consumption at the same point in the week rather than day to day — weekly allowances reset on a fixed schedule tied to your account, so same-day comparisons mislead. If the shape has not changed, the work is genuinely large and the honest answer is a bigger plan.