After Two Failed Attempts at Multi-Agent Context Sharing, Three Principles for Verified State

New_Difficulty_8152 · reddit · 2026-09-01

The author tackled the old problem of sharing state across multiple agents in a session—not conversation history, but verified state: what changed, what's blocked, and what evidence exists that a task is done.

Failed attempts

Maintaining handoff notes manually in Markdown: updates got skipped when tired, the next agent read stale context as current, and the human became a routing bottleneck.
Letting agents write the notes: an agent could write "tests pass" as easily as actually running them—the next agent inherited a story about the work, not verified work.

Three principles of the final build

State in explicit fields, not paragraphs—an agent can't make unresolved work disappear by writing a nicer summary.
The implementing agent can't approve its own work; a separate reviewer inspects the result against the original goal.
Machine-checkable claims need evidence attached to a specific version: "tests pass" is a claim; a test result bound to the exact commit hash is evidence, and evidence doesn't transfer after code changes.

30 days of dogfooding: 4,172 PRs merged across 16 repos by one maintainer; coordination overhead stayed flat from 3 to 10 agents; stale-context bugs dropped to near zero.

Remaining issue: the reviewer still sometimes can't distinguish "the goal changed mid-task" from "the implementation is wrong"—mitigated by a goal-hash both agents reference, but the UX needs work. The project is open-source (link in comments).

Original post →

After Two Failed Attempts at Multi-Agent Context Sharing, Three Principles for Verified State

More from coding & agent