Agents
A five-agent system, made watchable.
Five agents read raw career notes for a single role, cluster them into themes, draft prose, send the drafts to a Critic, revise, and emit the final structured highlight list the resume page consumes. Pre-recorded, deterministic, watchable in under ninety seconds.
Assembling: Zendesk Senior Staff Engineer, Technical Lead
Replay mode · pre-recorded fixture · no API calls
Five agents read raw career notes for a single role, cluster them into themes, draft prose, send the drafts to a Critic, revise twice, and emit the final structured highlight list the /resume page consumes.
MCTS framing
selection · expansion · simulation · backup
Selection
Pick the work item, gather raw inputs.
Expansion
Open the problem up. Cluster events into themes.
Simulation
Draft a candidate, propose it for review.
Backup
The Critic reviews and revises. Most production orchestrators omit this.
The Critic loop in the Backup column is the phase most production orchestrators omit. It is why the artifact does not read like every other resume.
Forward arrows are the assembly path. The dashed Critic to Editor edge is the MCTS backup phase.
Press play to start the run.
System prompts
Each agent has a short system prompt. These render here as content; the replay runtime does not execute them. A future Anthropic runtime sends them verbatim.
Space toggles play and pause. Arrow keys nudge the scrubber. R resets.
Why mock-first
The replay is canned because the goal here is the story, not the inference. A bit-for-bit reproducible run is the only way to polish to "watchable in ninety seconds." ADR 0004 documents the swap path: the visualization above never knows which runtime is behind the curtain.
A sandbox mode with the same cast against mocked streaming responses lands in Phase 7. A real Anthropic runtime lands behind a rate-limited server route in a later pass, gated by environment variable and token budget.