AI Orchestrator
DE Early Access

How it works

Workflow

One idea, routed end to end — and continued from a verified state.

  1. Idea Something to build or fix.
  2. Strategy Direction and priority.
  3. Decision Logged, not lost.
  4. Scoped task Routable and reviewable.
  5. Model routing Right tool and mode.
  6. Execution Within repo rules.
  7. Verification Reviewed and checked.
  8. Commit A known good state.
  9. Handoff Result preserved.
  10. Next action Clear continuation.
A representative path. The numbered order is the text alternative to the diagram.

Five layers

What each layer is responsible for.

  1. Strategy Layer

    Decides what matters and why — direction, priority and scope before any code is written.

  2. Project Memory Layer

    Holds context, decisions, tasks and status so the project continues across sessions.

  3. Execution Layer

    Carries out scoped tasks within repo and quality boundaries, on the routed model and mode.

  4. Verification Layer

    Reviews and verifies results before they are accepted — and decides when to escalate or stop.

  5. Growth Layer

    Feeds results back into SEO, quality and funnel workflows over time (v1.5+).

Each phase, explained

Open any phase for the detail.

The diagram above already shows the full path; the early steps (idea, strategy, decision) are yours, the rest are routed to and carried out by AI within your rules. Expand a phase for what actually happens — no expanding is required to follow the flow.

Idea

A raw need — something to build or fix. It is reconciled with the recorded project context before anything moves.

Strategy

Direction and priority are set first. The smallest change that serves the goal beats a large speculative one.

Decision

The choice is appended to the decision log, never buried in a chat. Later routing and scope can refer back to it.

Scoped task

One task with named allowed files, protected areas and a definition of done — so it is routable and reviewable.

Model routing

The task is matched to a tool, model and the smallest reasoning mode that can safely solve it — not the most expensive by default.

Execution

The change is made within the repo gate and worktree rules, so parallel work never overwrites itself.

Verification

Results are reviewed and checked — for risky work, by a role distinct from the one that did the work.

Commit

Only verified, in-scope files are staged. A commit marks a known-good state, never a blind “add everything”.

Handoff

The agent handoff and the project handoff are written together, so both layers stay in sync across sessions and models.

Next action

Project status is updated and exactly one next best action is set. The loop is ready to run again.