In 30 seconds
- Agents work on tasks with some autonomy, not just line-by-line suggestions.
- Humans set goals, constraints and review results.
- Clear specs, tests and guardrails are what make it work.
How it works
- A human writes a clear task with acceptance criteria.
- An agent plans, edits code and runs tests.
- The agent proposes a change.
- A human reviews, adjusts and merges, or sends it back.
What humans still own
- What to build and why: the backlog and priorities
- Specs and acceptance criteria
- Review and the decision to ship
- Architecture and cross-cutting concerns
Guardrails
- Strong automated tests as the agent's safety net
- Limited permissions: no direct production access
- WIP limits on agent output so review keeps up
- Logs of what each agent did
Myth vs reality
Myth · Agents are autonomous team members.
Reality · They act within limits humans set and check.
Myth · More agents means more output.
Reality · Output is capped by human review capacity.
Myth · Agents remove the need for tests.
Reality · Tests are what make agent work trustworthy.
FAQ
Which tasks suit agents best?
Well-specified, testable tasks with clear boundaries.
Do agents count toward velocity?
Count delivered value, whoever or whatever built it.
Who is accountable for agent-written code?
The human who approves and merges it.