CutCtx guides
How to Set a Context Budget for a Multi-Step Agent Task
A worksheet for budgeting task instructions, evidence, tool results, and response space without assuming a model's context limit or promising savings.
A context budget is a planning tool for deciding what information a multi-step task needs and how much room to reserve for the response. It helps teams notice when repeated instructions, long tool outputs, or stale material are crowding out evidence that matters to the next decision.
There is no universal token allocation that fits every model or task. Check the current model documentation and measure with the actual tokenizer and workflow before relying on a limit. The worksheet below is hypothetical and makes no claim about CutCtx performance.
List what each step needs to know
Write down the task goal, constraints, source evidence, current state, and output format. For a multi-step process, distinguish information needed by every step from information needed only at one decision point. This makes it easier to avoid carrying full intermediate output forward when a concise source-linked summary is enough.
Before shortening anything, preserve the facts and references needed to verify the next action. If a detail cannot be recovered later and could change the result, do not remove it solely to save space.
- Instructions and invariants that must remain visible.
- Evidence needed for the next decision, with source identifiers.
- Tool outputs that can be filtered, summarized, or retrieved again.
- Response and reasoning room needed to complete the task.
- A contingency reserve for retries or unexpected output.
Use a worksheet, then measure the actual request
Estimate categories separately before combining them. Then check the actual prompt or request using the provider's current counting method. Token counts depend on the model and representation; character counts and rough word counts are not reliable substitutes for a measured request size.
The token allocations in this example are planning assumptions. They are not recommendations, a supported model limit, or an estimate of achievable compression. Replace them with measurements for the task and model you actually use.
| Category | Planned tokens (hypothetical) | What belongs here |
|---|---|---|
| Instructions and fixed rules | 1,200 | Task, boundaries, required output |
| Selected evidence | 3,000 | Relevant source excerpts and identifiers |
| Current state and tool results | 2,000 | Only current findings needed for the next action |
| Response room | 1,500 | Expected answer or structured output |
| Contingency reserve | 800 | Retry or unexpected tool output |
| Planning total | 8,500 | Sum of these invented allocations |
Choose what to retain, summarize, or retrieve
Retain exact identifiers, constraints, decisions, and evidence that must be acted on. Summarize repetitive or intermediate material only when the summary preserves the distinctions relevant to the next step. Retrieve a source again when the exact text is important and retrieval is available and permitted.
A summary is a lossy transformation unless proven otherwise. Keep links or stable source references and spot-check summaries against the underlying material, especially for negation, dates, quantities, exceptions, and ownership.
| Information | Default treatment | Check before reducing |
|---|---|---|
| Task constraints and approval boundaries | Retain explicitly | Could a later step mistake what is allowed? |
| Source facts used in a decision | Retain with source pointer | Can a reviewer retrieve and verify the source? |
| Repeated boilerplate | Consider one canonical copy | Does any step depend on its exact wording? |
| Large tool output | Filter or summarize with provenance | Were errors, exceptions, or negative results dropped? |
| Unneeded intermediate output | Discard when safe | Could it be needed to reproduce or debug the result? |
Review the budget after a complete task
Record the measured input size, number of steps, tool-output sizes, retries, and whether the task met its quality criteria. If a workflow repeatedly exceeds its budget, identify whether the cause is duplicated instructions, broad retrieval, verbose tools, or an overlarge task. Change one factor at a time where practical.
Treat cost and quality as separate measures. A smaller prompt is not an improvement if it omits required evidence, increases rework, or makes a consequential answer harder to review.