I work from the command line, delegating tasks to sub-agents that run in parallel: implementation, debugging and documentation happening concurrently instead of serially. This note is the operating model, condensed.
The loop
Context and plan. Give each agent clear goals, relevant context, constraints, API assumptions, and an explicit completion contract.
Orchestrate and build. Break work into bounded tasks, connect the right tools, coordinate implementation across the stack.
Verify and debug. Inspect outputs, compare payloads, run checks, trace failures, iterate until acceptance criteria pass.
A fourth step sits outside the loop and gates every cycle: human review stays at every consequential decision, and tests, trade-offs and handover notes get written down.
What parallelism is actually for
Not typing speed. Waiting.
While one session runs a build or a test suite, another is drafting the migration, and a third is writing the PR description. The dead time in software delivery was never the typing. It was the gaps between the typing.
The human becomes the scheduler and the reviewer, which is the job that was always worth the salary.
Boundaries that keep it sane
One session per bounded task, not one session per project. Sessions that sprawl accumulate stale context and drift, and the drift is not obvious from inside the session.
Durable knowledge goes in the repo. CLAUDE.md, specs, docs. Never rely on a
session remembering anything.
Every merge is human-gated. The commit trail carries the AI co-author trailers, so the split of labour is auditable rather than a matter of vibes.