Decision Flowcharts

Is the task completable by a single agent in one context window?
├── Yes → Single agent (Level 0)
│   └── Does self-evaluation produce reliable quality signals?
│       ├── Yes (tests exist) → Single agent + deterministic eval
│       └── No (subjective) → Add separated evaluator (Level 1)
└── No → Multi-agent required
    ├── Are subtasks genuinely parallelizable?
    │   ├── Yes → Fan-Out/Gather topology
    │   └── No → Pipeline or Hierarchical
    ├── Is output verifiable automatically?
    │   └── Yes → Cascade (single first, escalate on failure)
    └── Does cost justify 15x token premium?
        ├── No → Decompose into sequential single-agent tasks
        └── Yes → Select topology from phase-02 table

What is the task duration?
├── Minutes (1-15 tool calls) → ReAct
├── Minutes-hours, quality-critical → Generator-Critic (cap 3 iterations)
├── Hours, mechanical/verifiable → Ralph Loop (fresh context per iteration)
├── Hours, unpredictable → Magentic-One Dual-Loop
└── Hours, parallelizable → Orchestrator-Worker

Expected task duration?
├── < 30 min → No management needed
├── 30 min - 3 hours → FIC at phase boundaries (target 40-60%)
├── 3+ hours → FIC + sub-agents OR one-session-per-task resets
└── Automated loop → One-session-per-task with external state

Is the output objectively testable?
├── Yes
│   ├── Test suite exists → Deterministic eval (no LLM needed)
│   └── No test suite → LLM-generated tests → deterministic eval
└── No (subjective quality)
    ├── Consistent domain → Fixed calibrated rubric with separated evaluator
    └── Variable domain → Adaptive rubric (AdaRubric)

Should I decompose this task?
├── Changes < 4 files AND fits in one session? → No decomposition
├── Natural independence boundaries exist? → Feature-level decomposition
├── Features have internal complexity? → Sprint-level with contracts
├── Complex dependencies? → DAG-based task graphs
└── Optimal granularity uncertain? → ADaPT (decompose on failure only)

What type of error?
├── Transient infrastructure → Exponential backoff (max 3 retries: 5s/15s/60s)
├── Logic error in generated code → Run tests, feed error → flush context if stuck
├── Doom loop detected → Revert code, flush context, restart with lessons only
├── Premature completion → QA agent validates against feature list
├── Multi-agent cascade → Circuit breaker + SagaLLM compensation
└── Context exhaustion → Checkpoint, start new session with summary

Should I keep this harness component?
├── Is it infrastructure (state, persistence, error recovery)?
│   └── Yes → Keep; re-evaluate implementation, not existence
├── Is it reasoning scaffolding?
│   ├── Has the model improved since this was added?
│   │   ├── No → Keep; schedule re-evaluation at next model release
│   │   └── Yes → Run 10 representative tasks without it
│   │       ├── Success rate holds → Remove permanently
│   │       ├── Degrades → Keep; document specific failure cases
│   │       └── Mixed → Implement conditional activation
│   └── Can you name exactly what breaks without it?
│       ├── No → Remove (Removal Test failed)
│       └── Yes → Keep if it passes Complexity Justification Matrix