Running coding agents on two machines exposed the real problem: the control plane
Competitive-Bend-143 · reddit · 2026-07-25
Running coding agents across machines turns the control plane into the real bottleneck
The author says months of running Claude Code and Codex on a Mac plus a Windows box taught them that the hard part is not the agents themselves, but the orchestration around them.
Key lessons from the setup:
- State visibility matters more than logs. A single screen showing every session state — generating, waiting for approval, stalled, idle — reduced constant checking.
- Approvals need to be one-tap. Phone push approvals were much smoother than hopping back to the terminal.
- Unattended work must be machine-checkable. The workflow uses isolated git worktrees, with the repo’s own typechecks and tests deciding what merges.
- Second opinions should be fanned out. The same read-only question is sent to replicas across machines/models, then disagreement is used as the signal.
The post argues that scaling beyond one machine makes the setup heavier, but the control-plane layer is what determines whether multi-agent work is actually usable.
More from coding & agent
- ChatGPT Work agent can now log into websites and keep sessions across runs — OpenAIDevs · 2026-07-25
- Practical multi-agent orchestration for Codex splits work into scout, worker, and coordinator roles — pvncher · 2026-07-25
- Internal CS research agent is already running with GPT-5.6 and Claude model choices — BenBajarin · 2026-07-25
- LiteParse 2.8.0 drops ImageMagick and speeds up image-to-PDF conversion up to 7.2× — llama_index · 2026-07-25
- Claude Opus 5 flips between 1/3 and 2/5 before finally settling on an answer — Sauers_ · 2026-07-25
- Opus 5 shines in async long-horizon work, with four agent patterns for real systems — RLanceMartin · 2026-07-25