AI Workflow Patterns.
A canonical directory of reusable architectural patterns for deterministic AI execution. These patterns provide the foundational logic for scaling high-fidelity workflows.
The Validation Loop
A recursive self-correction pattern where an Auditor Agent validates the output of an Executor Agent before final delivery.
- Schema checks required
- Recursive Error Correction
The Multi-Agent Swarm
Parallel execution of specialized nodes (Research, Synthesis, Formatting) managed by a central Orchestrator.
- Massive Parallelization
- Task-specific checks
Dynamic Context Injection
An optimization pattern that prunes and injects context based on the current execution node's specific needs.
- Measure token usage
- Increased Logic Density
The State Transition Map
A pattern that tracks session state across multi-turn interactions, allowing for 'recovery' from logic failures.
- State Persistence
- Failure Recovery
Architecture Reference Poster
Download the high-resolution AI Workflow Patterns reference poster for your workspace or dev team.
Internal reference · Illustrative example
Choose a pattern using a failure case
A customer-support draft can follow a bounded sequence: retrieve approved policy text, draft a response, verify its claims against that text, then send it for review. If retrieval finds no relevant policy, stop with an unresolved status. Generating another confident draft cannot repair missing evidence.
Verification checklist
- Test retrieval returning no documents.
- Limit correction attempts and record each failure.
- Prevent repeated sends when a run is retried.
- Compare a simple sequence with a more complex design on the same cases.
Adding agents or parallel steps creates coordination work and does not automatically improve accuracy. Select the smallest pattern that passes the acceptance cases. Measure time and resource use in the deployed environment rather than borrowing percentages from unrelated examples.
Compare fixed and exploratory control flow