Skip to content

End-to-End Orchestration

End-to-end orchestration is the delivery path that moves work from issue creation through enrichment, planning, implementation, review, and rework. It is a sequential pipeline with explicit handoffs between stages, not a single monolithic agent session. The current model delegates stages across GitHub, Ikidna, Coder, and a human reviewer; the Orchestrator primitive is the runtime component that operationalises it maintaining the task registry, owning dispatch decisions, and holding the live log stream from whichever agent is executing.

flowchart LR
    I[Issue] --> E[Ticket Enrichment]
    E -->|workable?| P[Planning]
    E -->|too big| D[Decompose into subissues]
    D --> E
    P --> G1{Plan gate}
    G1 -->|approved| X[Implementation]
    G1 -->|revise| P
    X --> V[Verification]
    V -->|fail| X
    V -->|pass| PR[Pull Request]
    PR --> G2{Review gate}
    G2 -->|changes requested| X
    G2 -->|approved| M[Merge]
    M --> O[Operate / observe]
    O -.feedback.-> I

The two gates plan approval before compute is spent and PR review before merge bracket the autonomous middle and concentrate human judgement where it is cheapest and highest-leverage (see Planning and Execution).

The proven unit of autonomous delivery, converged on across the field (Copilot, Jules, Factory, Devin), is one issue producing one reviewable PR, executed asynchronously. The pull request is the human-in-the-loop boundary. Standardising the output envelope matters as much as the work itself: every completed unit returns a PR with a plain-language summary of what changed, the test runs that prove it, and links back to the plan and issue. A predictable envelope is what makes review fast enough to keep up.

The review bottleneck is the central scaling constraint of autonomous delivery: as the system writes more, the limiting resource shifts from writing to reviewing. Investment in making changes easy to review clean diffs, an autofix loop that grinds CI/lint to green before review, summaries that explain intent is investment in throughput.

The pipeline is intentionally recursive. Review feedback can re-enter the path, re-trigger execution, and update the same work artifact. Failed or incomplete work generates new issues rather than silently completing. This makes Orchestration Triggers and durable state management critical for correctness a recovery trigger must distinguish a genuine completion from a stalled run that never produced its artifact.

Stages run in sequence; parallelism is bounded to the safe cases (read-only fan-out within enrichment and review, surface-partitioned independent implementation) per the Coordination Model. The Orchestrator merges parallel branches only after each passes verification, so conflicting implicit decisions cannot reach main.