Autonomy direction
One agent or many?
A multi-agent environment is not automatically smarter. It is a coordination design: who plans, who specializes, what information crosses boundaries, and which actions require approval.
One context, one owner
Good when the task is coherent and the agent needs to see the whole problem. Fewer handoff failures and simpler debugging.
Central coordination
A manager keeps the conversation and asks focused specialists for research, coding, testing, or review. The manager assembles the result.
Explicit ownership transfer
A specialist takes over a phase or conversation. Useful when a new context is cleaner, but boundaries and state must be designed.
A safe autonomy ladder
Observe
Read files, inspect state, and produce a plan. No external side effects.
Suggest
Draft commands, patches, messages, or decisions for a human to review.
Act reversibly
Make bounded edits, run tests, or create artifacts with a clear diff and undo path.
Coordinate
Delegate narrow tasks in parallel, then require a synthesis and evidence check.
Operate
Allow recurring or external actions only with scope, budgets, logs, approval gates, and a stop mechanism.
Design the interfaces between agents
| Interface | Make explicit | Good handoff result |
|---|---|---|
| Task | Objective, scope, exclusions, success criteria | A specialist knows what done means |
| Context | Relevant files, facts, assumptions, unknowns | No need to rediscover the entire project |
| Output | Format, evidence, changed paths, open questions | The next agent can verify and continue |
| Authority | Read, write, network, spending, or messaging permissions | Side effects happen only inside the intended boundary |
| Failure | Timeout, retry, escalation, and human stop behavior | Partial work is visible instead of silently lost |
Good first experiments
Parallel review
Ask separate reviewers to inspect accessibility, security, and test coverage. Give each the same commit or snapshot, then compare findings and resolve conflicts centrally.
Research then build
Let one agent collect current primary sources and another turn only those sources into an implementation outline. Keep the source packet visible.
Generator then critic
Have one agent draft an artifact and a second check it against a fixed rubric. The critic should report evidence and failures, not merely a confidence score.
Watch with a stop button
For recurring work, begin with read-only monitoring and meaningful-change alerts. Add mutation only after false positives, retries, and costs are understood.
When not to add another agent
If the task is unclear, the repository is unmapped, the tools are unsafe, or the output cannot be evaluated, another agent adds more context and more failure surfaces. Improve the brief, interfaces, and proof first.