Skip to content

patterns.md

Latest commit

 

History

History
54 lines (42 loc) · 3.51 KB

patterns.md

File metadata and controls

54 lines (42 loc) · 3.51 KB
Raw

Workflow patterns

Distilled from Anthropic's "A harness for every task: dynamic workflows in Claude Code" (Thariq Shihipar & Sid Bidasaria) and adapted to the flow API. Runnable versions are in examples/patterns/.

The three failure modes workflows fight

Failure mode What it is Pattern that fights it
Agentic laziness the model declares a complex job done after partial progress loop-until-done, generate-and-filter
Self-preferential bias the model prefers its own output when grading it adversarial verification, tournament (separate judge)
Goal drift the original objective decays across many turns / compaction classify-and-act (decide the route once, up front)

The patterns

  • classify-and-act — a cheap classifier decides the task type, then route to the right handler/tier. → examples/patterns/classify_and_act.py
  • fan-out-and-synthesize — split into independent steps, each with its own clean context, then merge. Use when steps must not cross-contaminate. → examples/audit_template.py
  • adversarial verification — for each producing agent, spawn a separate agent to refute its output against a rubric. Default to "refuted" when unsure. → audit_template
  • generate-and-filter — generate many candidates, dedupe, filter by a rubric/verifier, keep only survivors. → examples/patterns/generate_and_filter.py
  • tournament / comparative judgment — N approaches compete; judges pick winners by pairwise comparison (more reliable than absolute scoring). → examples/patterns/tournament.py
  • loop-until-done — for unknown-size work, loop until K dry rounds (no new results), bounded by budget — not a fixed count. → examples/patterns/loop_until_done.py

Verification (deeper)

  • deep verification — one agent extracts factual claims; a subagent checks each claim; another rates source quality.
  • multi-rule — one verifier agent per rule; a "skeptic" agent reviews the rules themselves to keep false-positives down.

When to use — and when not to

Use for long-running massively-parallel work, structured problems needing verification, adversarial/independent-hypothesis tasks, large migrations, deep research, qualitative sorting at 1000+ scale, root-cause investigation.

Don't for ordinary tasks. Ask: does it really need more compute? Workflows use significantly more tokens — most coding tasks do not need a panel of 5 reviewers. Always set a budget (--budget '{"max_usd": 2}' or max_tokens); flow enforces it as a hard ceiling and the router picks the cheapest capable model per leaf.

Security: quarantine

Leaves that read untrusted public content must not take high-privilege actions. flow's authored-script allowlist (authoring.validate) is a guardrail against the model, not a sandbox against an adversary — run untrusted scripts with executor_kind="process" + OS isolation.