Sync-Plan, Async-Execute¶
Zakariasson's personal operating rhythm for managing 5–10 parallel cloud agents. The clearest articulation in recent coverage of how a "manager of agents" actually spends their day.
The pattern¶
- Plan synchronously, in focus, using a model as sparring partner. Source from Notion/Slack/Linear. Produce a long spec or plan.
- Execute asynchronously — spawn cloud agent, hand it the full spec, let it run.
- While waiting (agents take minutes to hours): one synchronous task, or light scrolling, or another async spawn.
- When an agent returns: review output (often video of UI, not code), prompt corrections, merge if good.
Why front-load intent¶
Short prompts fail on long tasks. Agents need the full intent up front — spec, constraints, verification plan, success criteria — because mid-run corrections cost a round-trip of minutes.
"You kind of like frontload the context to the agents either through a plan or a long spec and then you send them off and then you let them go."
As models improve, prompt length shrinks — but intent clarity doesn't. The compression target is "what does 'done' look like?"
Concurrency sweet spot¶
5–10 agents at once, ~4 different repos/areas simultaneously. More and you can't hold the map in your head. parallel-agent-competitions is a subset of this — race N agents on one task, pick winner.
Context-switching cost¶
Zakariasson is explicit: "The sad reality in some sense is that there's going to be a lot of context switching." Front-end, back-end, database, docs, side projects, Twitter bug reports — all in flight. The manager mindset is context-switching.
Cross-references¶
- software-factory — this is how the manager runs it
- agent-as-junior-engineer — Hashimoto's parallel: async delegate with clear spec
- driving-into-mud — the failure mode this rhythm is trying to avoid (underspec → drift → waste)
- isolated-agent-vms — the substrate that makes async possible