Skip to content

Offload Thresholds by Payload Origin (Tool Output and Pasted Input)

Payload-origin offload thresholds decide when tool results and human messages become file references; Deep Agents documents different limits without testing the split.

Documented offload behavior

LangChain's October 14, 2026 Deep Agents post describes an offload policy keyed to who supplied the payload:

Payload origin Documented threshold Storage path
Tool result Roughly 20,000 tokens /large_tool_results/{tool_call_id}
Human message 50,000 tokens /conversation_history/

The harness saves the full content before it enters the conversation. The model receives the opening and closing lines plus a file path. It can retrieve further detail with read_file. These are the post's documented settings; the post reports no comparison of alternative thresholds (LangChain).

This admission policy differs from summarizing accumulated conversation history and retaining content by semantic type. Origin determines when a payload becomes a preview. It does not classify which statements within the payload matter.

What the coding study supports

Lewis evaluates a different mechanism: older tool results shrink as the context fills, while task text and other conversation messages stay unchanged. The treatment also includes stall detection, interventions, and command safeguards (study, sections 2–3).

On 169 SWE-bench Verified tasks, the study uses a 20,480-token window and a fixed 480-second attempt endpoint. Mean per-task fail-to-pass fraction rises from 28% to 49%; complete solutions rise from 43 to 72. These results belong to the combined treatment. The experiment does not isolate any component's effect (Lewis, abstract and section 3.1).

That evidence cannot validate the Deep Agents thresholds or establish a benefit from treating human messages differently. See half-life truncation for the age-based policy and its evaluation limits.

Reading an offload trace

A missing file read alone does not establish a wrong answer. LangChain describes retrieval as necessary when the agent needs more detail; the preview remains available (documented retrieval behavior). The implication for review is to check whether the answer requires information absent from that preview. A read count alone cannot answer that question.

The sources above provide no tested recipe for choosing your own pair of thresholds. Treat threshold tuning as a separate evaluation question, with task correctness as an outcome.

Key Takeaways

  • Record payload origin when examining an offload trace; the documented admission rules differ for tools and human messages.
  • Separate evidence for a combined harness change from evidence for a particular threshold.
  • Review the information needed for the answer before classifying an unread offloaded file as a failure.
Emerging — single recent source Last reviewed: 2026-10-28