---
slug: "planning-and-worker-flow"
name: "Planning and Worker Flow"
packType: "harness"
canonicalPattern: "hybrid"
version: "0.1.0"
trust: "Verified"
publisher: "Agent Workspace Labs"
updatedAt: "2026-04-15"
---

# Planning and Worker Flow

> Tiered orchestration for questions that are too broad for a single call.

## Summary

A multi-step harness pattern that uses a narrow planning pass, typed worker calls, and a constrained synthesis pass instead of letting one prompt attempt everything at once.

## Install

```sh
npx attrition-sh pack install planning-and-worker-flow
```

### Claude Code / AGENTS.md snippet

```md
# Planning and Worker Flow
# See: /packs/planning-and-worker-flow

```

## Contract

_No execution contract defined for this pack type._

## Layers

_No three-layer split defined for this pack type._

## Use When

- The user asks a broad or multi-entity question.
- You need cross-tool synthesis with traceable intermediate outputs.
- A single-shot answer tends to overgeneralize or hallucinate missing evidence.

## Avoid When

- A deterministic query or a single issue packet already contains everything needed.
- The overhead of planning exceeds the complexity of the task.

## Key Outcomes

- Broad questions are decomposed into smaller evidence-gathering steps.
- Worker outputs stay typed and auditable before synthesis.
- The final answer is grounded in a bounded set of prior results instead of hidden chain-of-thought.

## Minimal Instructions

Use a three-phase harness:
1. Plan the smallest useful set of steps.
2. Execute workers with typed outputs and explicit dependencies.
3. Synthesize only from validated worker results.

Do not let the model answer broad questions directly when multiple tools or scopes are involved.

## Full Instructions

You are implementing a harness for questions that exceed a safe single-call boundary.

Phase 1: Plan
- decide which tools or internal workers are required
- keep the plan narrow
- include dependencies and step purpose

Phase 2: Execute
- run steps by tier
- allow parallel workers only when outputs are independent
- persist success, failure, duration, and result summary

Phase 3: Synthesize
- read only validated worker outputs
- cite the evidence sources or record IDs used
- do not invent facts not present in worker results

Favor typed worker outputs over freeform intermediate prose.

## Evaluation Checklist

- Was planning narrower than the original question?
- Did each worker have a clear scope and output contract?
- Did synthesis only use worker results and cited evidence?
- Could the run recover or re-plan if a worker failed?

## Failure Modes

- **[MID] Planner creates too many steps with no dependency discipline.**
  - Trigger: (legacy — trigger not separated)
  - Prevention: (legacy — no explicit prevention)
- **[MID] Workers return unstructured prose that is hard to validate.**
  - Trigger: (legacy — trigger not separated)
  - Prevention: (legacy — no explicit prevention)
- **[MID] Synthesis ignores worker results and drifts back into generic model prose.**
  - Trigger: (legacy — trigger not separated)
  - Prevention: (legacy — no explicit prevention)

## Transfer Matrix

_No measured cross-model transfer data._

## Telemetry

_No telemetry recorded._

## Security Review

_No security review on file._

## Compares With

_No comparative data._

## Related Packs

_No related packs._

## Changelog

_No changelog entries._

## Sources

- [Anthropic: Building effective agents](https://www.anthropic.com/research/building-effective-agents) — Canonical workflow patterns such as prompt chaining, routing, parallelization, and evaluator loops.
- [LangGraph overview](https://docs.langchain.com/oss/python/langgraph) — Reference for controlled multi-step and multi-agent orchestration.
- [LangGraph product page](https://www.langchain.com/langgraph) — High-level framing for reliable agent orchestration.

## Examples

- [Builder preview workspace](/workspace-a)
- [Reviewer preview workspace](/workspace-b)