Ship Packets, Not Transcripts

Brief the agent with accepted intent, authority, relevant sources and required checks.

In the loop

Feedback from the last cycle becomes the next brief. Select the intent and evidence the agent needs, then identify their source owners in Factor II.

The rule

Give the agent a bounded brief: the accepted outcome, examples of success, scope, authority, relevant sources and required checks.

Include decisions already made, so a later worker does not contradict them. A brief points to the existing request or work item. It does not create a competing version of the intent.

This applies to a report, an investigation, an operational task or a code change. Retrieve history when it can change a decision; pass a cited finding instead of an entire transcript. Keep a line only if dropping it would change what the agent does. A clear request can be the brief without another document.

The evidencesupported

Studies measure what long inputs and standing instruction files do to a model. No study compares a curated brief with a full transcript handed to the next agent.

Chroma, 2025 evaluated 18 models and found performance grew less reliable as the input got longer. Du et al., 2025 found performance fell by 13.9% to 85% as inputs grew, even when the model could retrieve every relevant fact.
Gloaguen et al., 2026 ran coding agents with and without repository context files. The files had no significant effect on task success and added more than 20% to inference cost. With all other documentation removed, generated files improved success by 2.7%. Context helped when it carried what the agent could not find elsewhere.
Yang et al., 2025 found underspecified prompts twice as likely to regress when the model or the prompt changed. Listing every requirement did not consistently fix it, because requirements conflict and models have limited ability to follow instructions.

One dissent: Cognition, 2025 argues for sharing full agent traces, because a worker that lacks an earlier decision will make a conflicting one. A packet therefore carries the decisions already made along with the task. The length studies test single prompts; long agent sessions are less studied.

Put it to work

State what output the caller wants and what makes it acceptable.
Link the source request, relevant evidence and known constraints.
Name what the agent may change and what needs approval.
Carry unresolved questions forward; amend the source when intent changes.
Tell the agent to ask when the brief leaves a choice open.

Example · the monthly operations report

The brief asks for a draft monthly operations report from the approved ledger. Totals must reconcile, claims must cite sources and personal details must be excluded. Publishing needs operator approval. Last month's reconciliation failure belongs in the brief because it changes this month's checks.

Failure signal

The handoff copies a transcript but leaves success undefined.
A rewritten plan changes the request without caller agreement.
The agent cannot identify the sources behind its answer.

Done looks like

The agent can state the requested output, its boundaries and its checks from the caller's intent and a small set of cited sources.