One agent to coordinate the rest — auto-review and auto-implement
A new orchestrator runs risk review on every plan and, when you opt in, drafts the fix as a PR — so one human approval can drive an entire review-and-fix cycle.
Until this release, each InfraDots agent did one thing in isolation: review a plan, summarize a change, propose a fix. Useful — but stitching them together was still a human's job.
What's new
The agent orchestrator coordinates per-stage agents into a single workflow:
- On plan, the orchestrator runs the risk-review agent. Findings are categorized (critical, high, low, advisory) and posted to your VCS PR.
- Optionally, on review, the orchestrator hands the findings to the implement agent, which proposes a code change to address them. The change comes back to you as a PR commit you can accept, reject, or amend — no auto-merging.
The orchestrator decides which agents to call based on the workspace's configuration and the stage of the job, so you don't wire each one in by hand.
Why it matters
The most common workflow we see is: plan fires → reviewer flags an issue → engineer reads the flag → engineer writes the fix → engineer pushes → re-plan. That's four context switches for what is often a one-line adjustment. With the orchestrator + auto-implement, the same loop is one click: review the proposed fix, accept it, and the next plan is the one you ship.
Safety
Auto-implement is opt-in per workspace and never bypasses approval. Even when the implement agent drafts a fix, the resulting plan still goes through your normal review gates — the orchestrator just removes the typing in between.
How to use it
Risk review is on by default once an agent is attached to your workspace. Auto-implement is a separate toggle in the workspace's agent settings.