To shrink an agent’s context safely, prune tool-result payloads by explicit, deterministic rules—not by editing the reasoning or call structure needed to continue the run. Keep provider-native reasoning artifacts intact, preserve the sequence and identifiers that connect each function call to its result, and remove only results that are stale, duplicated, or explicitly outside the active allowlist.
What you can prune—and what must stay intact
Tool output can make multi-step agent contexts grow quickly: search results, file contents, browser extracts, and execution logs may be much larger than the calls that produced them. The safe boundary is data integrity. Reduce expendable payloads while retaining the context needed to replay or continue the run faithfully.
- Potentially prune: old tool-result payloads that are stale, duplicated, or explicitly excluded by policy, once no later step depends on them.
- Keep: assistant reasoning items and provider-native reasoning artifacts, function-call items and outputs, call identifiers, ordering, and arguments needed to associate a result with its request.
- Keep evidence that is still referenced: a result is not safe to remove merely because it is old if a later step relies on it.
OpenAI’s official reasoning guide recommends preserving items between the last user message and the function-call output untouched during truncation and optimization. When several functions run consecutively, it also recommends passing the reasoning items, function-call items, and function-call outputs together. Apply those replay rules rather than treating the conversation as ordinary text that can be freely shortened.
Apply a deterministic pruning policy
Define the pruning decision in code or configuration, not by asking a model to rewrite an active conversation. A model-generated summary can be useful for a separate archive or later context, but it is not a substitute for preserving the state and associations required to continue the current run.
- Tag each tool result. Record the tool name, call ID, timestamp, turn, and downstream references. Keep call arguments and ordering metadata so the result can be joined to the right request.
- Assign a rule class. Mark items to retain, summarize outside the active chain, or drop after a defined safe horizon. Make the rule explicit, such as retaining all results referenced by later steps.
- Protect the active sequence. Before pruning, retain the items from the latest user message through the matching function-call output unless the provider explicitly permits a different transformation.
- Preserve provider-native state. Do not edit or partially remove reasoning artifacts required by the provider: OpenAI reasoning items or encrypted content, Anthropic thinking blocks, and Gemini thought signatures.
- Log the decision. Record which rule was applied and the original result’s hash so the pruning decision can be audited. Retain raw history in durable storage where policy permits.
- Replay and verify. Check that each kept result remains adjacent to, or correctly associated with, its call, and that no later step has lost evidence it still needs.
Provider-specific replay requirements
| Provider or approach | What to preserve | Pruning implication |
|---|---|---|
| OpenAI | Reasoning items and the sequence around function calls. Reasoning tokens are not exposed as ordinary text. | Use documented reasoning-context and replay mechanisms; do not assume hidden reasoning can be treated as editable text. Keep consecutive reasoning items, calls, and outputs together. |
| Anthropic | Complete thinking blocks during tool use. | Anthropic’s documentation says: “Pass every thinking block back to the API complete and unmodified.” Do not partially edit a block; older thinking blocks may be filtered according to model policy. |
| Google Gemini | Thought signatures and their associated function-call context. | Google Cloud describes a thought signature as a “save state” for resuming after a function result. Preserve it consistently; partial context can degrade performance. |
| OpenClaw-style local pruning | Normal conversation text is not rewritten by session pruning; raw stored history and a replay view are separate concerns. | Tool-result trimming can be scoped with allow/deny lists. Older processed image blocks may be replaced in a replay view while raw history remains preserved. |
These are provider-specific semantics, not interchangeable formats. A policy that is safe for a local replay representation does not establish that the same transformation is valid for a provider’s reasoning artifacts.
Choose a pruning rule that preserves causality
For each candidate result, evaluate six things: whether it is eligible for removal; whether the decision is deterministic or model-generated; how the provider represents reasoning state; whether raw history is retained; the resulting token and latency change; and what happens if a required result is missing.
A useful default is to retain anything in the active call sequence or referenced by a later step, and consider pruning only results that are both outside that sequence and no longer needed under an explicit rule. If a result is needed but too large to keep in full, create a separate, clearly labeled summary for future use only when the workflow allows it; do not silently substitute that summary for required original state.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Measure savings without mistaking a benchmark for a guarantee
The Squeez research paper on arXiv (2026) reports 0.86 recall, 0.80 F1, and 92% input-token removal in its coding-agent evaluation. Those figures describe that paper’s evaluation, not a universal production guarantee. Your own token reduction, latency, and failure behavior depend on the tools, workloads, pruning rules, and provider replay requirements you use.
Recommended Free Tools
Test with representative multi-step runs. Verify that retained outputs still match their call IDs and that the agent can continue when a pruned result is absent. Treat a missing required result as a replay failure to surface and recover from—not as permission to invent or infer its contents.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
