Agents made typing cheap.
Engineering is what is left.

This is the argument behind AgentCrew, and the standard we hold ourselves to when we write about it. If you disagree with any of it, the blog is where we work it out in public.

1. The bottleneck moved from hands to judgment.

A coding agent can produce a plausible diff for almost anything in minutes. What it cannot do is decide what should be built, notice when the requirement changed halfway through, or tell you honestly whether the thing works. Those were always the hard parts of engineering. Agents did not remove them; they exposed them.

2. A prompt is not a specification.

Every serious engineering discipline writes down what it is about to do before doing it. Agents make that more necessary, not less, because they will cheerfully build the wrong thing at full speed. A requirement with a stable ID and acceptance criteria is the smallest unit of accountability we know of. AgentCrew refuses to run a plan that does not cover every one of them.

3. "Done" is a claim until there is evidence.

An agent's completion message is the least reliable artefact in the whole pipeline. It is generated by the same process that made the mistake. We hold that a change is complete only when an independent verification run has produced evidence tied to the exact commit and the current requirements, and that the moment either changes, the evidence is stale. This is not distrust of agents. It is the same standard we apply to ourselves in code review.

4. Isolation is kindness.

Two agents editing one working tree is a merge conflict with extra steps. Every task should own its worktree, its branch and its session, and integration should be an explicit, recorded merge. When that merge conflicts, a human should decide. Automatic conflict resolution is automatic damage.

5. Humans approve; they should not babysit.

The right number of interruptions is the number of decisions only you can make. For an engineering change that is usually two: is this the right plan, and is this ready to deliver. Everything else the crew should either do within authority you already granted, or record and route to an inbox you read on your schedule.

6. Recovery is a feature, not an incident.

Laptops sleep. Quotas run out. Processes die. A tool that requires a ritual to recover from ordinary life is not finished. State must be durable, work must be checkpointed, and a restart should resume, not restart.

7. Your context belongs to you.

Project knowledge, decisions, evidence and plans should live in files and databases you can open, export and delete. A tool that makes your own understanding of your codebase inaccessible outside it has taken something from you. AgentCrew keeps everything local and readable.

8. Use the fewest agents that work.

Swarms are a demo, not a method. More agents mean more coordination, more quota, and more places for the requirement to drift. Plan the smallest graph that covers the work, run what can genuinely run in parallel, and stop.

For solo founders

You are already the product manager, the reviewer and the on-call engineer. The last thing you need is to also be the message bus between three agent windows. Set the outcome, approve the plan, read the evidence. That is the whole job, and it is the job you wanted.

For small product teams

The risk with agents in a team is not that they are bad. It is that everyone uses them differently, and the pull requests stop meaning the same thing. One crew, one evidence standard and one draft-PR handoff give the team a shared definition of done again, whichever runtime each engineer prefers.

For engineers across many repositories

Re-explaining the codebase to every new session is the tax nobody budgets for. Imported work keeps its source links, project knowledge stays attributable and versioned, and the Bridge carries context between projects so the crew arrives already briefed.