Build and test an agent

The Agent Workspace is available in both the browser app and local Workstation. It combines behavior, a model connection, Context Lab knowledge and typed capabilities into a versioned agent. The browser and desktop use the same agent factory; cloud-specific availability matters for local models and external executors.

From starter to first conversation

1. Choose a starting point on the home screen or + New agent inside a project. Describe one clear job. A draft is private to the project until you publish it.
2. Open Model. In the browser, configure OpenAI with your API key and model ID, or use clearly labelled Rehearsal without an AI call. The desktop also supports installed CLIs, Ollama and compatible endpoints. Check connection verifies it before you rely on a model response.
3. In Knowledge, write a note or add a text document. See what the agent would receive resolves a sample message without calling the model. Use and Stop using control this agent's access. The same knowledge can be reused by another agent only when that agent is explicitly allowed to see it.
4. Open Try it and ask a representative question. Read What happened in this turn: draft or release, model, delivered knowledge, available capabilities and proposed calls. The model's answer is a separate observation from the context trace.

A change to behavior, knowledge, capability access or model connection can make prior evaluations stale. Save the draft and run its scenarios again before release.

Give the agent one capability

Under Capabilities → Add a capability, describe what the agent may request and declare typed input fields. Decide whether it only reads information or changes data. The interface distinguishes contract defined, allowed for this agent, executor configured, credential available, exact contract checked and real test call succeeded. One badge does not imply the others.

The browser can define and authorize a contract, but it cannot connect an arbitrary external executor on the shared cloud host. Open the project locally to connect an HTTP API or a service implementing the SEIF executor contract, verify compatibility and make a real test call. Data-changing calls need an operator to review their exact arguments. A model or application credential cannot approve its own action. Your system must still enforce user authorization and idempotency.

Test and publish

Create scenarios in Tests with realistic messages and expectations: what the reply must mention or omit, which actions it may propose and latency limits. Run tests checks structural context and then runs the model or labelled rehearsal. A passing evaluation belongs to that exact saved draft and environment; generated answers can vary on another run. A rehearsal pass is not evidence of real-model quality.

Create release from this draft requires current passing evidence. Publishing points new conversations in one environment to that release; applications keep the stable URL. A staged share requires a baseline, and existing sessions keep their original release. Production requires a real model evaluation and live context connections. Integrate gives the project-scoped application credential and a server-side request example. Report outcomes from your application to make Performance useful.

Use browser instructions, connection guide, and releases and Context Builds for the other half of the workflow.