Use the Agent

Ask Tako’s in-browser coding agent to plan, write, execute, and verify a reproducible Tako Script while you retain control over scope, scientific assumptions, and deliverables.

The Agent uses your connected Codex account. The generated script and atomistic calculations run in Tako’s browser workers; model requests necessarily travel through the connected Codex service.

Connect your account

Open Agent from the top toolbar. Choose one supported connection route:

• Sign in with ChatGPT through the device-code flow.
• Load auth.json from the Codex CLI credential file.
• Paste auth.json into the connection form.

Imported credentials are stored in the current browser profile. Treat them as sensitive: do not include them in screenshots, workspace exports, scripts, or reports. Clear the credential from Agent settings on a shared machine.

Start a task

Write a prompt that specifies:

1. The scientific question.
2. Inputs or structures to use.
3. Required method or the criteria for choosing one.
4. Validation gates.
5. Exact deliverable paths.
6. Claims that must not be made.

Example:

Map a coarse water PES over O–H distance and H–O–H angle with GFN2. Write reference/run.js, output/summary.json, a contour SVG, an interactive Plotly surface, and analysis.md. Report this as a demonstration, not a transition-state calculation.

Choose model/reasoning effort, then click Run. Higher effort can help with ambiguous multi-stage design.

Read the plan before execution

The first accepted turn must write a plan containing ## Deliverables and a checklist. Review it for scope, method, validation, and filenames. A well-formed plan can still contain a scientifically poor choice; steer before expensive calculations begin.

The Agent verifies declared deliverable paths are non-empty before it can report success. This checks existence, not scientific validity.

Steer, stop, and resume

While a run is active, the input becomes Steer. Send corrections such as a missing comparison site, wrong model domain, or stronger convergence gate. Steering messages are consumed at turn boundaries and should be reflected in the plan.

Use Stop when continuing would waste resources or produce invalid work. A stopped conversation remains inspectable. Use New only when a clean context is preferable to a follow-up.

Inspect the deliverables

Do not accept only the final chat message. Decide whether the run is reproducible and scientifically supportable, not merely complete according to its file-existence gate.

Review order

1. Read the original request and final plan.md for declared scope and checklist state.
2. Confirm every deliverable exists and has the intended role.
3. Read reference/run.js completely as the executable method.
4. Verify input provenance, atom counts, cells, charges, and constraints.
5. Inspect each calculation log and quality gate, including convergence diagnostics.
6. Recalculate headline values from output/summary.json and supporting tables.
7. Compare plots with their source data and labels.
8. Read claims and limitations in analysis.md.

Audit table

Verify independently

Where risk warrants, rerun the script, alter one parameter to test sensitivity, compare against a known small case, or validate with another method. Agent self-review is useful but not independent evidence.

Context, compaction, and history

The context meter reports active usage for the selected model. Compaction preserves the original request, recent turns, and a history bridge that points back to plan/log/files. The plan and workspace artifacts—not conversational memory—should remain the durable state.

Terminal runs are stored in a capped local run history with prompt, model, effort, status, usage, duration, summary, files, and log tail. Export history when you need an audit record; it is not a substitute for the scientific workspace.

Privacy boundary

Agent conversation/context leaves the browser through your Codex connection. Tako Script and calculation execution remain client-side, apart from downloading static runtimes/model weights. Review what structures, reports, screenshots, or files are included in prompts/tool context before using sensitive data.

Troubleshooting