The single-agent ceiling
Stuff a single AI agent with too many tools and it starts to forget the goal mid-flow. The fix that\'s become standard in 2026 is the planner / worker / critic trio.
Roles
- Planner (Opus 4.7) — reads the request, produces a step-by-step plan, never touches tools.
- Worker (Sonnet 4.6 or Haiku 4.5) — executes one step at a time with a small, fixed tool set.
- Critic (Sonnet 4.6) — reviews each completed step against the plan, flags drift, requests retries.
Why it works
Each agent gets a narrow context — its own role, its own tools, its own success criteria. That means smaller prompts, faster runs, fewer hallucinations and a clear audit trail you can show a client.
A reference build
For a research-report generator, the trio looked like this:
- Planner produces a 7-step research outline.
- Worker hits search, reads sources, drafts each section.
- Critic checks for unsupported claims and missing citations.
- Planner approves the final assembly.
Total cost dropped roughly 35% versus a single-agent version, and human-edit time on the output dropped by half.