orchestration-patterns

The Orchestration Patterns skill provides a robust architectural framework for managing complex, multi-step agentic workflows. Designed for development teams using the SAFe (Scaled Agile Framework) methodology, it codifies how AI agents should progress through long-running sessions, ensure verifiable outcomes, and handle escalations. By implementing Simon Willison's iterative agent loop, this skill guides models to break down tasks into goal-oriented cycles—execute, validate, and escalate—ensuring no work is completed without objective evidence. This is particularly useful for engineers and delivery managers coordinating subagents across diverse development lifecycles.

• Implements a rigorous evidence-based delivery model requiring test results, screenshots, and logs before considering a task complete.

• Enforces a mandatory QAS (Quality Assurance Specialist) gate for all PRs to ensure independent verification and bias-free code review.

• Provides standardized session checkpointing patterns to prevent context loss during long-running work or handoffs between agents.

• Includes structured escalation paths for Technical Delivery Managers (TDM) when agents face blockers, architectural ambiguity, or cross-team dependencies.

• Integrates with external tools like Linear, GitHub, and CI/CD pipelines to maintain an audit trail for all software delivery actions.

• Invoke this skill when orchestrating multi-step implementations or coordinating multiple specialized subagents (e.g., BE, FE, and Data engineers).

• Use it to generate structured evidence reports before merge, ensuring all commit messages and documentation follow project-specific patterns.

• When blocked, utilize the provided escalation templates to feed the TDM agent clear, actionable context and historical attempts.

• Expected inputs include task acceptance criteria, while expected outputs are validated technical implementations, finalized QA reports, and updated tracking tickets.

• Constraints include a mandatory adherence to the iterative execution loop; agents must validate findings via CI commands before proceeding to the next step, avoiding the 'trust me, it works' pitfall.