Skip to content
⚠️ Before You Start
Risk: Medium
Duration: 2-3 hours
Last Validated: Not yet validated (draft)
Kind: Decision guide
Risk Note: Medium (over-engineered planners raise cost and failure modes)

Choose a planner pattern: workflow, ReAct, plan-and-execute, reflexion or orchestrator-worker

Draft outline: not validated

Drafted with AI from official sources on 16 Oct 2026. Do not use on a live system until it is written, dry-run by someone other than the author and signed off (planned 2026-12-11). Author: James · dry-run: Sanjay Sarvesh · sign-off: Prasanna. Work items: AGT-03-T01, AGT-03-T02, AGT-03-T03, AGT-03-T04.

1. Summary

During solution design, or when an agent is too slow, too costly or too unreliable and the planner is suspected.

Risk: Medium (over-engineered planners raise cost and failure modes). Typical duration: 2-3 hours.

2. When to use and when not to

Use it when: During solution design, or when an agent is too slow, too costly or too unreliable and the planner is suspected.

Do not use it when: to be written by the author.

3. What you must know first

  • Workflows have fixed paths; agents choose their own steps
  • ReAct interleaves reasoning and tool actions
  • Plan-and-execute separates a planning call from execution steps
  • Reflexion uses self-critique and memory to retry
  • Each extra planning or critique call adds cost and latency

4. Options

Outline of the decision process; the author describes each option.

  1. Classify the task: fixed steps, dependent steps, independent subtasks or checkable output
  2. Start from the cheapest pattern in the decision table that fits
  3. Estimate model calls per task for the chosen pattern
  4. Prototype on ten examples and measure success, calls and latency
  5. Move one level up only if the quality gap is shown
  6. Write the expected autonomy and limits for the client
  7. Record the decision YAML

Draft snippet: Planner decision table and record

# docs/decisions/planner-pattern.yaml
use_case: ${USE_CASE}
inputs:
  steps_known_in_advance: false
  max_tools_per_task: 6
  latency_budget_seconds: 20
  cost_budget_per_task: low
  quality_gate_available: true
decision_table:
  - when: one tool call or a fixed sequence answers it
    choose: deterministic workflow (prompt chaining or routing)
  - when: few tools, short tasks, steps depend on tool results
    choose: ReAct tool loop with step budget
  - when: long tasks with known sub-goals and a need to show the plan
    choose: plan-and-execute with bounded replanning
  - when: an automatic check can score the output
    choose: evaluator-optimizer or reflexion loop with max iterations
  - when: subtasks are independent
    choose: orchestrator-worker with parallel fan-out
chosen: ReAct tool loop with step budget
client_expectation_note: autonomy is bounded; humans approve side effects

Snippets to add:

  • snippets/docs/decisions/planner-pattern.yaml: Planner decision record (yaml)
  • snippets/docs/client/autonomy-statement.md: Client autonomy statement template (markdown)

5. Decision table

Situation Choose Why
to be written

To be written by the author from the outline above.

7. Known issues and fixes

Symptom Likely cause Fix
Plan-and-execute slower than expected Planning call plus replans on every task Use ReAct for short tasks; bound replans
Reflexion loop never converges No reliable evaluator Use a checkable criterion or drop reflexion
Client expects fully autonomous handling Autonomy not discussed early Share the limits section of the decision record at design sign-off

8. Monitoring and alerts

Signal Alert when Action
Model calls per task in production versus design estimate More than 2x the estimate weekly Revisit the planner choice

9. AWS and open-source equivalents

The patterns are framework-neutral. AgentCore and OpenAI Agents SDK run the same loops; LangGraph documents prompt chaining, routing, orchestrator-worker and evaluator-optimizer patterns.

10. Validation record

Pending. Planned dry-run scenario: A second engineer applies the guide to three sample tasks and must choose the same patterns as the agent lead, with call estimates within 30 percent of a measured prototype.

11. Sources