FIELD NOTE / X
Parallel agents need shared decisions.
The short film, the complete written thought, and the evidence behind it.
Parallel agents need shared decisions.
Short video caption
Five agents need more than one giant prompt. Isolate worktrees, name one owner per decision, push sourced context deltas, and block merge on unresolved contradictions. Measure missed constraints and accepted work, not message volume. Illustrative. #EricFieldNotes
Full written post / accessible read
Parallel agents are powerful because they can investigate, code and test at once. But give five agents one repo and a giant context dump and you may get five versions of why an auth migration is safe. More tokens do not create one decision owner.
Agent A changes the schema. Agent B tests an old branch. Agent C remembers a failed tenant case, but that fact never reaches the worker shipping the patch. The conflict looks like a merge problem; it is a coordination problem.
Give each worker a worktree and a named task. Let specialist agents keep narrow durable memory of prior decisions and exceptions. A supervisor sends a small, sourced delta to the affected worker and records its acknowledgement.
Before merge, require one decision record, the latest specialist messages, acknowledgement, and an independent test of the disputed boundary. Do this because agent-to-agent messaging matters only when it changes the decision, not when it adds another chat.
#EricFieldNotes
Evidence and boundary
On-screen boundary: ILLUSTRATIVE TEAM RUN. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.