FIELD NOTE / LINKEDIN
Design shared-state control before parallel agent work.
The short film, the complete written thought, and the evidence behind it.
The LinkedIn conversation link will follow its public release.
Design shared-state control before parallel agent work.
Video caption
Design shared-state control before parallel agent work.
Scope and handoff must survive its session.
Branch passes do not compose automatically.
My rule: Watch integration debt, not agent activity.
#EricFieldNotes
Full written post / accessibility read
Worktrees are the first rule for a personal agent fleet, but not the last. Two agents can change different files and still implement incompatible interpretations of a product decision. A useful fleet needs role ownership, source-backed messages, durable state and a trusted merge gate.
Assign one worktree, allowed files, acceptance fixture and escalation condition. The agent records source, change, evidence, uncertainty and owner in a durable packet. A supervising agent can monitor progress and relay new facts without flooding every worker's context, but it must not invent authority.
Suppose one branch changes API entitlements and another changes a default flag. Both local suites pass. On the merged artifact, run the protected customer-decision fixture and a deliberate negative control. Bind status to that merge digest and have one integrator approve the product interaction.
My rule: worktrees isolate edits, sourced messages align assumptions, and a separately controlled merge gate validates the combined build. Do this because parallel agents raise throughput only when shared decisions and product outcomes survive integration.
#EricFieldNotes
Four-beat scene transcript
1. Design shared-state control before parallel agent work.
Worktrees are the first rule for a personal agent fleet, but not the last. Two agents can change different files and still implement incompatible interpretations of a product decision. A useful fleet needs role ownership, source-backed messages, durable state and a trusted merge gate.
Visual: Concurrency moves conflict from files to decisions.
2. Give each agent a bounded contract.
Assign one worktree, allowed files, acceptance fixture and escalation condition. The agent records source, change, evidence, uncertainty and owner in a durable packet. A supervising agent can monitor progress and relay new facts without flooding every worker's context, but it must not invent authority.
Visual: Scope and handoff must survive its session.
3. Re-evaluate at merge.
Suppose one branch changes API entitlements and another changes a default flag. Both local suites pass. On the merged artifact, run the protected customer-decision fixture and a deliberate negative control. Bind status to that merge digest and have one integrator approve the product interaction.
Visual: Branch passes do not compose automatically.
4. Scale by accepted combinations.
My rule: worktrees isolate edits, sourced messages align assumptions, and a separately controlled merge gate validates the combined build. Do this because parallel agents raise throughput only when shared decisions and product outcomes survive integration.
Visual: Watch integration debt, not agent activity.
Research and claim limits
- OpenAI Agents SDK/Codex orchestration example (S07)
- GitHub Docs: ruleset required status checks (S146)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.