Why separate stages can help
Separate sessions and typed artifacts can reduce context drift and make each stage's work inspectable.
A single-agent loop keeps one conversation around the task. That shape can create three recurring costs:
Confirmation bias. The same agent may review its own reasoning and assumptions.
Context drift. As a conversation grows, relevant context can be harder to manage.
The babysitting tax. You may still have to steer, correct, and re-verify a stream of edits.
The artifact boundary
In the full multiagent path, Niki splits the work into Planner, Coder, Tester and Reviewer, separate sessions that exchange typed artifacts: a TaskSpec, a unified diff, a test report, and a verdict. Each artifact is JSON-schema validated before handoff. In auto mode, a low-complexity task can collapse to Planner plus a solo Coder.
The Tester receives the spec and Coder artifact rather than a shared conversation. The Reviewer receives those artifacts plus the Tester evidence. Separate sessions narrow context contamination; they do not make a stage blind to the evidence it needs.
What the code records
This isn't philosophy: isolation records mirror the wiring, Red agents receive evidence-only projections, and shared-model review raises warnings. Those mechanisms narrow the problem; they do not prove correctness. See the claims audit for code-level details.
A completed run may produce a local niki/<id> branch and changes.patch; it does not create a GitHub pull request.