Overview
AI agents — reasoning models that can observe, plan and act across a session — introduce a new layer into system workflows. The experiment explores where an agent can genuinely assist: not as a passive lookup tool, but as something that holds context, notices patterns, and proposes the next move.
Early work runs inside the Observatory build itself, using Claude as a development collaborator on the same codebase it helps reason about. The interesting constraint is that the agent's view of a system is always mediated — it reads files, not environments — so the experiment is also about making system state legible enough for an agent to operate on.
What's being tested
- Whether an agent can maintain useful context across long build sessions (architecture, intent, constraints) without requiring constant re-briefing
- How to structure system descriptions so an agent can navigate and reason about them
- Where agent judgment is genuinely additive vs. where human judgment is non-negotiable
Status
Actively used in the Observatory build. Findings will surface as practices stabilise.