1,500 Agent Sessions Show Tool I/O Eats 57% of Context While Dialogue Is Just 2%
rajistics · x · 2026-11-01
Analyzing 1,500 OpenHands sessions (averaging 78k tokens each), jdpruettt breaks down where a long-context agent's window actually goes: tool execution results 37%, system prompts and tool descriptions 23% (15k-18k tokens just explaining tools and environment), tool call arguments 20%, user messages 9%, and direct model responses only 2%.
Roughly 57% of context is raw tool I/O, meaning curating tool outputs matters far more than trimming dialogue. rajistics confirms the same pattern in his own analysis across Agent Canvas and Codex: agents mostly read files, write patches, and retrieve info.
More from coding & agent
- Blogger lets Grok spend up to $100 autonomously and receives a LEGO Harry Potter gift — venturetwins · 2026-11-01
- TensorAgent Adds Local Multimodal Indexing for Documents, Images, Audio and Video — fuzhongkai · 2026-11-01
- Theo claims new model hallucinates 10x worse in real codebases; bugs traced to ACP integration — cgarciae88 · 2026-11-01
- One prompt to an AI agent adds note-taking and search to X bookmarks — HaijunXia · 2026-11-01
- Running 16 desktop AI agents at once on a single Mac, clicking via accessibility refs — cramforce · 2026-11-01
- Grok helps user find 74GB Cursor chat backup, doubles free disk space — prasenx · 2026-11-01