Definition
A wave of agents is a coordination pattern where you dispatch many small, focused agents in parallel (and then in subsequent waves) so each agent owns one bounded unit of work.
It’s “parallel work with controlled merge pressure.”
Why waves (not just “all at once”)
Waves exist because real systems have constraints:
- rate limits (API, CI minutes, model quotas),
- merge conflicts,
- shared context (one agent changes docs, another changes the same file),
- and human review bandwidth.
Waves let you batch work into sets that can be validated and integrated safely.
A simple operating model
- Wave 0 (map): one explore pass to group issues and identify dependencies.
- Wave 1 (quick wins): low-conflict, high-signal fixes.
- Wave 2 (risky changes): security, workflows, refactors—more gates.
- Wave 3+ (deep work): things with real investigation costs.
Each wave should have a defined “merge condition” (gates + review).
When it works well
- Backlogs with many independent issues
- Repetitive fixes (docs consistency, small refactors)
- Cleanup work with strong automation gates
When it goes poorly
- Work items are tightly coupled (you get conflict storms)
- There are no shared conventions (agents “solve” the same problem differently)
- There is no integration role (agents produce diffs, nobody ships)
Guardrails
- Use a gated multi-agent flow for every wave.
- Require build receipts from every agent.
- Have an explicit coordinator role (this is a core AgOps responsibility).
Practical rule
One wave, one merge condition. If you can’t state the merge condition, the wave isn’t defined.