Quick Start

Concrete techniques from the method. Each links to the full treatment. For the behavioral patterns behind many of these techniques, see agent patterns.

Directing agents

Push back on “needs human judgment.” Agents anchor to what’s in their current context and escalate questions their tools could resolve.

  • Before accepting an escalation, check whether the agent could answer its own question from available data — version history, dependency patterns, usage spread.
  • Most of the time, directed investigation resolves it.
  • → Calibration

Scope agent context to what the task needs. Unnecessary context creates anchoring; insufficient context creates blind spots.

  • A reviewer evaluating code quality shouldn’t see the implementation plan — it shifts evaluation from what the code does to whether it matches intent.
  • → Context scoping

Use wrong answers productively. When something feels off but you can’t articulate it, ask the agent to characterize the problem.

  • The agent’s attempt gives that inarticulate sense something concrete to push against.
  • The correction often crystallizes a sharper distinction than either side started with.
  • → Productive friction

Invite the agent to push back. Explicitly ask the agent to counter intuitions, find weaknesses in the direction, or challenge assumptions. The exchange is more productive when bidirectional, and the human controls when to open that space. → Friction by invitation

Establish halt authority. Tell agents they can and should stop if the task is misconceived — “if this approach cannot work, say so and explain why rather than producing a partial result.” Agents are biased toward completing tasks even when the premise is wrong. → Scope completion bias

Running parallel agents

Parallelize by default. When two investigations don’t depend on each other, run them simultaneously. Each agent works one area with bounded context; synthesis happens at the orchestration level. → Parallel execution

Layer agents for large investigations. One agent per area of concern, plus a broader agent looking at cross-cutting themes. The cross-cutting agent catches dependencies and shared patterns that area-specific agents miss. → Parallel investigation

Planning

Have agents surface their assumptions. Agents fill ambiguity with assumptions silently. Tell them to surface the assumptions they’re making — investigate the ones that matter and catch the ones that are wrong before they’re baked into the work. → What a plan contains

Review

Use fresh agents for later review rounds. After many revision cycles, agents treat the artifact’s current shape as fixed — producing fewer structural suggestions even when structural changes are warranted. Spin up fresh reviewer agents for later rounds to counteract this. → Convergence in review

Hide the implementation plan from evaluative reviewers. Reviewers checking code quality should see the code and the problem being solved, not the design decisions. The plan creates anchoring — evaluation shifts from what the code does to whether it matches intent.

  • Alignment reviews (“did we build what we planned?”) are different — those need the plan.
  • → Review

Documentation

Course-correct on structure early. Structural changes before editorial polish are cheap; after, they waste the polish.

  • Options: propose a structure, give the agent constraints and let it determine structure, or give it a free hand.
  • When an instinct says a structure doesn’t fit but the reason isn’t clear, ask the agent to evaluate the choice against the content — it won’t reliably volunteer objections, but it’s good at evaluating a specific proposal when asked.
  • → Draft and shape

Batch feedback. Accumulate observations across a draft and deliver them together so the agent sees the full pattern, not individual corrections. This produces better revisions than piecemeal feedback. → Batch feedback

Research

Track findings and assumptions as shared artifacts. Findings and assumptions serve as context passed between agents and across context boundaries.

  • Provenance matters — agents can’t reconstruct what produced a finding once it leaves their context.
  • Both need triage: investigate what’s load-bearing, defer what’s non-blocking.
  • → What accumulates

Condense to keep agents focused. As findings and assumptions accumulate, agents spend more effort processing noise than doing useful work.

  • Periodically condense — simplify, reduce noise, highlight the larger overarching concerns — so agents operate on a cleaner signal.
  • Preserve the reasoning behind each finding; a fresh agent needs that to work with the material.
  • → Knowledge accumulation

Knowing when to stop

Watch for convergence. New passes confirm rather than extend. Findings stop surprising. Findings shift from structural to minor. When this holds across several perspectives, the work is approaching completion. → Convergence

Watch for false convergence. Fatigue feels like convergence — the reviewer is tired, not the artifact complete.

  • Agents have their own version: after many rounds they anchor to the current shape.
  • If a fresh perspective finds significant issues, the work continues regardless of how many prior rounds came back clean.
This site uses Just the Docs, a documentation theme for Jekyll.