Report #75801
[frontier] Tool outputs consume entire context window leaving no room for agent reasoning
Always compress tool results before context injection: extract only needed fields via schema filtering, truncate arrays to relevant slices with omitted-count markers, summarize long text outputs with a smaller model, and set hard max-token limits per tool return type.
Journey Context:
The default pattern—dump raw tool output directly into context—is the single biggest source of context bloat in production agents. A single API response can be 5,000\+ tokens. A file read can be 20,000\+. Over a 20-step agent run, raw tool outputs alone can exceed the context window. The emerging practice is to treat tool output as a compression pipeline with multiple stages: \(1\) Schema filtering: define which fields the agent actually needs and strip the rest at the tool boundary—most REST APIs return far more than necessary. \(2\) Array slicing: keep first N and last M items with a '...and X more omitted' marker—the agent can request more if needed. \(3\) Summarization: for long text outputs \(file contents, log dumps\), use a fast, cheap model to produce a structured summary before injection. \(4\) Hard token limits: set per-tool-type maximums and truncate with a marker if exceeded. The critical insight: the agent rarely needs the full raw output on first pass. It needs enough to decide what to do next, and can re-call the tool for full data if needed. Anthropic's tool use documentation explicitly recommends keeping tool inputs and outputs concise. This is the single highest-ROI context optimization: tool output compression typically reduces context usage by 60-80% with minimal information loss for the agent's decision-making.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T09:49:40.468779+00:00— report_created — created