# The agent The Neurarch agent designs and edits architectures by emitting **structured actions** against the typed graph (add a component, wire a connection, update parameters, scale a dimension), never by rewriting code as free text. That constraint is what makes the safety story real: structured actions can be checked; arbitrary diffs cannot. ## The guardrail pipeline Before any agent action touches your model, it runs through six gates: 1. **Impact**: the blast radius of the change (downstream layers, shape-changing, weight-carrying). 2. **Parameter explosion**: a rough total-params estimate; warns past a configurable threshold, blocks at twice the threshold. 3. **Cycle**: would the proposed connections close a cycle? 4. **Orphan**: does a new component end up connected to nothing? 5. **Shape**: sub-millisecond shape inference diffs the sandboxed result against the baseline; structural breakage blocks the edit. 6. **Parameter range**: per-field validation (a dropout above 1, a stride of 0, a head count of 0). Findings render in an impact preview before anything applies. If you cancel a preview, the agent is told why on the next turn so it adjusts rather than re-emitting the same plan. An optional plan-coherence check (an LLM judging whether the plan matches your request) can be enabled in Settings. ## What it can do Build a full architecture from a task description, complete a partial canvas, fix flagged findings, apply bulk edits ("halve the width of every FFN"), explain the current design, and read training curves to suggest targeted changes. Destructive actions are gated and previewed; the completion flows never delete or replace layers you placed yourself. ## Practicalities - The agent requires sign-in (15 agent messages/day free; more on paid tiers). Design, import, checks, and most exports need no account. - Hosted agent calls route through our backend proxy. You can instead bring your own provider API key; it stays in your browser and is never sent to our servers. - Auto-tier routing (when using Claude) classifies each prompt and picks the cheapest adequate model; your selected model acts as a ceiling. For what the agent is allowed to see and store, read .