# AI Agent Workshop ## Learning goal Give a simulated workflow a goal. Follow its plan, tool result and review, then introduce failure to see where it must stop. ## Model and limits An explicit local finite-state teaching workflow. There is no language model, autonomous external action or network tool. Results come from a mock calculator, catalog or checklist. The review applies simple scripted rules and is not a real factual verification. Retry limits, evidence requirements, step budgets and checkpoint decisions are visible. ## Primary references - [Anthropic: building effective agents](https://www.anthropic.com/research/building-effective-agents) ## Experiment design - **A successful workflow**: The tool returns evidence and the local review passes. Parameters: `{}`. - **Recover from an error**: The first tool call fails; the workflow retries. Parameters: `{"failures":1}`. - **Missing evidence**: Review fails, requests revision and eventually blocks. Parameters: `{"missing":1}`. - **Human checkpoint**: Inspect the returned result before approving or rejecting. Parameters: `{"human":1}`. - **Tool unavailable**: The required tool is disabled; no observation is invented. Parameters: `{"tools":0}`. These presets are design predictions to investigate, not claims from a usability study. All readouts, animation positions and exported observations come from the same local model. The 2D view uses the same state. Saved setups restart from their parameters, rather than claiming to resume exact progress. Captured baselines are stored separately and survive resets and preset changes. ## Verification Run `node --test apps/shared/discovery/tests/models.test.cjs` for numerical invariants, limiting cases, determinism and failure-state checks. Run `node apps/shared/tests/catalog.test.cjs` for routes and references. Browser and keyboard validation must be reported separately from numerical tests; automated tests are not user research. ## Discovery studio interaction layer The studio adds selectable model objects, ray-based 3D picking, keyboard-accessible object selection, an inspector, scene editing where appropriate, model-specific stepping, and 24 measured challenges across the twelve labs. This is an interaction redesign, not a claim of measured learning improvement. Most parameter changes now preserve current state. Structural changes explicitly start a new run; the exact per-app list is in `apps/shared/discovery/studio-model.js`. Current settings, independently captured baselines and challenge progress retain separate storage. Undo/redo keeps up to 30 edit checkpoints; the timeline keeps the latest 80 model snapshots. Resuming from a historical state creates a new branch. Saved setups still restart from parameters; they do not claim to retain the timeline or local particle pin edits. Additional instruments derive from the existing model and cited references. Camera and light sweeps hold the remaining inputs fixed; neural curves separate training and validation; compression uses actual coefficients and recomputed quality samples; traffic uses recorded car positions/speeds; factory queues and utilisation use actual events and busy time. Visible instruments can continue collecting data while the hidden 3D view stops drawing. Run `node --test apps/shared/discovery/tests/studio.test.cjs` for live edits, preserved state, all challenge solutions, direct manipulation, rewind, picking, instrument exports and controller integration. These DOM/renderer-double checks do not replace rendered browser and device review. ## Detailed 3D apparatus Shared procedural geometry in `apps/shared/discovery/apparatus.js` adds mechanical housings, lens and iris parts, cables, wheels, connectors, moving machine heads and subject-specific markings. The scene uses bevelled housings, smoother curved geometry, instanced material families, a key/fill/rim light arrangement and bounded soft shadows. Cloth has a procedural weave bump map attached to its original model coordinates. Full detail can be switched to Essential detail without resetting the model. All movements use model state or model time: wheel markers use road position, machine heads use scheduled job progress, conveyors use accumulated worker time, iris blades use the illustrated aperture, and network packets use their scheduled flights. Signal and energy markers illustrate direction, not measured transit speed. The fan is an indicator of the wind setting; it does not add a fluid simulation. Shells, grippers, cables and surface textures do not change the underlying collision model. The 2D fallback retains the same state and geometry. Additional checks: `node --test apps/shared/discovery/tests/apparatus.test.cjs`. Geometry, state preservation, bounded batching, semantic selection, pause/rewind consistency and fallback behavior are checked with renderer doubles; browser pixels and device frame rates remain unverified. ## Authored model assets The physical-object labs now load original Blender-authored GLB assets through `model-assets.js` and `model-worlds.js`: articulated industrial arm, mirrorless camera, car (overview and inspection versions), network router, gantry station, optical laser and companion character. The geometry, authored material surfaces and packed texture maps replace corresponding primitive bodies only after loading succeeds. Source `.blend` files, reproducible Python generators, manifests and inspected Blender renders are retained in `assets/models/` and `scripts/`. Mathematical data surfaces stay tied to their original calculations. The robot uses actual solved joints; camera iris hinges respond to the aperture and a Cutaway toggle exposes the sensor/rays; car wheels and brake lights follow model position/acceleration; factory heads use scheduled job progress. GLB equipment covers remain illustrative, not new collision or optical physics. Camera cutaway distances are display distances; numerical thin-lens readings remain authoritative. The companion uses the motion model's translation and squash/stretch. Assets load on demand, repeated equipment shares instanced geometry/materials, and only the selected close-up car upgrades from the lighter road mesh. Failure preserves the diagram and exposes Retry model. Inspect model fits the selected geometry rather than applying a fixed camera distance. Full/Essential detail, pause, rewind and existing models/data remain available. A procedural studio reflection environment lights PBR surfaces without external network requests. Run `node --test apps/shared/discovery/tests/model-assets.test.cjs` for real GLB parsing, file loading, texture/UV presence, hierarchy transforms, exact solver alignment, cutaways, pause/rewind, fitting and car detail selection. Asset product renders were visually inspected in Blender. Browser pixels and device frame rates are still unverified; product renders are not screenshots of the web app. ## Bespoke rebuild (2026-10) The page no longer uses the shared discovery template or a 3D scene: an agent workflow is a graph and a sequence, so two purpose-built 2D diagrams explain it better than depth would. - `index.html`, `styles.css`: new page (question cards, live readouts, controls column beside the stage, step budget and run controls directly under the diagram, answer panel, trace, three checked challenges, model notes, limits, sources). On phones the controls sit straight under the diagram and a small sticky status bar (state, step count, steps needed, Step) appears once the diagram scrolls away. - `agent-derive.js`: fixed labels and colour groups for each model state, the mock-tool call strings shown in the trace, and `forecast()`, a dry run of the unchanged model with the same settings (checkpoints auto-approved) that gives the "needs N steps" marker. - `views.js`: **Loop map**, a hand-laid state diagram (landscape on desktop, portrait on phones) showing the current state, the path taken, visit counts, switched-off gates and the arrow into the stop box; **Timeline**, workflow step across and state down, so retries and revisions show as zig-zags and the step limit as a wall. - `app.js`: controller. All states, counts, results and evidence come from `model.js` (`create`, `advance`, `next`, `action`); `model.js` and `config.js` are unchanged. Limits: the "thoughts" in the trace are the model's fixed transition messages; the tool error text is a label for a simulated failure; the review checks only whether evidence is attached. The Revisions readout counts revise loops actually taken (the model's internal counter reaches 2 at the moment it blocks). Approval at a checkpoint is remembered by the model, so a revision after approval goes straight to review. Animation speed is display pacing only.