agent-workflows

Agent Workflows

Experimental multi-agent workflow patterns for Claude Code. Twenty-one skills that orchestrate parallel agents to research, brainstorm, prototype, debate, stress-test, plan, execute, and review — with Beads for cross-session task tracking. See docs/ for detailed reference guides.

Skills

Ideation & Research

brainstorm

Structured ideation with shape-uniqueness enforcement. Researches prior art, then generates N ideas where each one is structurally distinct from all others. Use when exploring solution spaces, designing features, or making architecture decisions.

diverge

Multi-perspective research. Spawns N independent agents with different lenses (technical, UX, risk) to explore a question. Synthesizes findings into a PRD. Use before building, when you need to understand trade-offs.

fracture

Competitive problem decomposition. Spawns N agents each proposing a different decomposition of a problem — by user journey, data flow, failure domain, team boundary, deployment unit, API surface, or lifecycle. Reveals how framing shapes the solution space.

Decision & Debate

converge

Structured debate. Takes divergent findings and runs moderated rounds where agents advocate positions, challenge each other, and propose compromises. Use when resolving tensions or making architectural decisions.

constraint-inversion

What-if constraint removal. Takes a design with stated constraints and spawns N agents, each removing one constraint to explore what becomes possible. Reveals which constraints are truly load-bearing vs assumed.

Risk & Validation

premortem

Prospective failure analysis. Spawns N agents that each write a narrative from the future where the project failed, each for a different root cause. Synthesizes into a risk registry with severity ratings and mitigations. Based on Klein's premortem technique.

stress-test

Parallel adversarial analysis. Spawns N agents each tasked with breaking a design or codebase from a different attack vector (security, scale, concurrency, edge cases, dependencies, data integrity, operations). Produces a vulnerability map with severity ratings and fixes.

diffuse

Blast radius impact mapping. Takes a proposed change (diff, migration, dependency upgrade) and spawns agents tracing impact through call graph, data flow, test coverage, deployment pipeline, and user-facing behavior. Produces a combined blast radius map.

Prototyping & Recombination

diverge-prototype

Parallel prototyping. Spawns N agents in isolated git worktrees to build different implementations of a PRD, then compares results. Use for architecture spikes and implementation exploration.

crossbreed

Structural recombination. Takes 2-3 existing prototypes and spawns agents to create hybrid designs, each with a different dominant parent. Documents grafting decisions and seam risks.

entangle

Dependent co-design of coupled systems. Spawns one agent per subsystem in separate worktrees, sharing a single evolving interface contract. A coordinator merges contract proposals and broadcasts updates. Agents iterate until the contract stabilizes.

migrate

Parallel migration strategy exploration. Takes an old-to-new system transition and spawns agents in worktrees, each prototyping a different migration strategy (big-bang, strangler fig, parallel run, feature flags). Compares rollback safety, downtime, and data integrity.

Specification & Verification

contract

Specification generation from examples. Takes API calls, test cases, or I/O pairs and spawns N agents to independently infer the specification. Where agents agree, the spec is clear; where they diverge, the examples are ambiguous.

replicate

Independent verification by reimplementation. Spawns N agents in isolated worktrees to implement the same spec independently. Convergence reveals clear spec; divergence reveals ambiguity. The divergence map is the output.

Planning & Analysis

distill

Progressive compression. Runs an artifact through a chain of agents, each compressing by 50%. The key insight: what gets dropped at each layer reveals the priority hierarchy. The waste product is the signal.

Debugging

bisect

Binary search for root cause. Uses agents to perform binary search through git history, configuration space, or dependency versions to isolate a regression. Adaptive topology: each step depends on the previous result.

Execution

focus