LLM Orchestrator: Make agent-assisted engineering accountable
CreatorPublic project · 2026
An open-source orchestration framework for breaking substantial engineering work into scoped tasks, coordinating agents and verifying results before completion.
Context & constraints
Giving several agents work is easy; keeping their responsibilities, review and completion criteria clear is harder. LLM Orchestrator makes those engineering controls explicit. The public project documents task classification, planning, agent scopes, parallel execution, phase gates and verification, with adapters for different coding-agent environments.
- Work needs an explicit scope and completion criteria before it can be handed to an agent.
- Parallel execution introduces coordination and integration concerns, even when individual tasks look independent.
- Different coding-agent environments have different interfaces and capabilities.
- An agent’s report of completion needs to connect to review and verification evidence.
My contribution
- Created an open-source framework documenting how to classify engineering tasks, form a plan and assign bounded work.
- Made task scope, phase gates, review and verification part of the orchestration workflow.
- Documented parallel-agent execution and the need to coordinate results into a reviewed delivery.
- Provided adapter concepts and instructions for Claude Code, Codex, OpenCode and Kilo.
- Published the project so the workflow and engineering approach can be inspected directly.
Technology context Agent orchestration · Claude Code · Codex · OpenCode · Kilo · Verification workflows
Orchestration workflow
A conceptual rendering of the public repository’s workflow: classify and plan work, assign scoped tasks through an adapter, then review and verify results.
Components & responsibilities
- Engineering task
- Problem, constraints and intended outcome.
- Classification & planning
- Choose an approach and divide work into explicit scopes.
- Environment adapter
- Connect the workflow to a coding-agent environment.
- Scoped agent work
- Bounded execution, including parallel tasks where appropriate.
- Review & verification gates
- Check results against scope and completion criteria.
- Accountable result
- A result connected to its review and verification.
Connections & flow
- Engineering task to Classification & planningUnderstand & scope
- Classification & planning to Environment adapterExecution plan
- Environment adapter to Scoped agent workDispatch scoped work
- Scoped agent work to Review & verification gatesResults & evidence
- Review & verification gates to Scoped agent workRequired corrections
- Review & verification gates to Accountable resultCompletion criteria met
Conceptual view of the project’s documented workflow, including human review.
Decisions & trade-offs
Make scope explicit before dispatch
A bounded task is easier to review and integrate than an open-ended request.
Trade-off: Planning and scope definition add work before execution begins.
Treat completion as a verification gate
A successful-sounding response is insufficient evidence that a change meets its requirements.
Trade-off: Review and verification use time and context, so they need to be proportionate to the task.
Separate the workflow from environment adapters
Common engineering principles can be reused across different agent tools.
Trade-off: Adapters must account for each environment’s behavior and limits.
Outcome & operation
Published an inspectable orchestration framework that makes scope, coordination, review and verification explicit. It is a concrete example of my work on responsible AI-assisted software delivery.
Notes & references
- Public repository
- The source and workflow documentation are available on GitHub.