budgeting-agentic-work
Skill0 starsChanged 49 days agoMITView on GitHub
Use when launching work that will spend real money or context at scale — multi-agent passes, multi-phase implementations, long-running pipelines — when a session approaches a budget or context ceiling, and when unplanned rework cost has appeared.
In this repository18 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- addressing-review-feedbackSkill
- delegating-to-subagentsSkill
- git-syncSkill
- implementation-gatesSkill
- implement-planSkill
- landing-merged-workSkill
- maintaining-project-memorySkill
- plan-featureSkill
- reporting-defects-upstreamSkill
- review-prSkill
- tracking-open-asksSkill
- verifying-before-sendingSkill
- writing-code-commentsSkill
- writing-commit-messagesSkill
- writing-for-audiencesSkill
- writing-plan-docsSkill
- writing-pr-descriptionsSkill
The file
Files
Budgeting Agentic Work
Agentic work spends the operator's money, and the bill arrives after the decisions that ran it up. Three failure shapes, each from real spend: a budget wall hit mid-phase, losing paid work in interrupted agents; a rework round silently doubling a pass's cost; and compute left running after its purpose was served. Neighbors own adjacent ground — delegating-to-subagents owns mid-run rescue budgets for silent agents, maintaining-project-memory owns the session-end starter prompt — this skill owns how the spend is structured and disclosed.
Phase the work to survive interruption
Money and context are both meters that can hit their ceiling mid-run; structure expensive work so that:
- Each phase ends durable. Results committed or written to disk, in a state a fresh session can resume from, before the next phase spends anything. An interruption then costs at most the current phase — never the run.
- The split is planned before the wall forces it. Watch the context meter the way the operator watches the money meter; a session past roughly three-quarters of its window plans its ending at a phase boundary instead of discovering it mid-agent. The boundary ends with the literal next-session starter prompt —
maintaining-project-memoryowns its form.
Price the pass before running it
Before an expensive fan-out or long pass, state in one line what will run and what it is expected to cost — scale, agent count, or a money estimate where the harness shows one — so the operator can redirect before the spending starts, not after. review-pr's axis triage is this rule applied to reviews; the rule holds for any expensive pass. Proportionality is the test: verification and enrichment spend scales to what it protects, and a pass expected to cost more than the work it checks runs only on the operator's explicit yes.
Rework cost is a process defect
When a pass's bill balloons — one real implementation round's cost more than doubled after a post-hoc review forced rework — the extra spend has a cause, and the cause is usually a check that ran after the expensive pass instead of before it. When rework cost appears: name the cause, move the check earlier in the process so the next run doesn't pay it, and report the incident plainly. Never present unplanned spend as routine — an unexplained cost spike is the operator's failure signal, and "that's just what it cost" converts a process defect into a recurring bill.
Stop what you started
A server, watcher, agent, or scheduled job whose purpose is served gets stopped — proactively, not on request. Anything left running needs an articulated reason the operator has seen. Two costs, not one: idle compute is spend without a consumer, and a forgotten process corrupts later work — a stale server answering probes fakes verification results long after everyone forgot it was there. (delegating-to-subagents owns deciding whether a silent agent is dead or working; this rule is for after the work is done.)
Red flags
| Thought | Reality |
| "Run it all in one go, it's simpler" | One ceiling hit mid-run loses everything unfinished. Phase it, durable ends |
| "The context will probably last" | Plan the split before the wall picks the boundary for you |
| "The operator will see the cost in the bill" | The bill arrives after the decisions. Price the pass before it runs |
| "A deep verification pass is always worth it" | Spend scales to stakes. Costlier than the work it checks → operator's call |
| "The rework is done, no need to dwell on it" | Unexplained spend recurs. Name the cause, move the check earlier, report it |
| "I'll leave the server up in case we need it" | An articulated reason or it stops — idle compute spends and fakes results |
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

