What is an agent?
An agent uses a language model to choose actions, read their results, and respond to a request. In Amodal, those actions are tools: querying an API, reading a document, writing a store record, or running a custom handler.
Model
The model decides what to do with the information available in the conversation. Amodal sends it instructions, messages, and the tools the session is allowed to use.
Choose the model in provider settings. Model selection is separate from the agent's source files.
Runtime loop
For each user message, the runtime:
- Sends the conversation and available tools to the model.
- Streams the model's response to the client.
- Executes requested tools, applying the checks supported by each execution path.
- Adds tool results to the conversation and calls the model again.
- Stops when the model finishes, the user cancels, an error occurs, or a configured limit is reached.
A question that needs no tools can finish after one model call. Longer tasks can require several calls. See The Core Loop.
Agent source
Your files define what the agent should do and what it can access:
| Files | Purpose |
|---|---|
amodal.json | Project identity and runtime features |
agents/<name>/ | Instructions and declared resources for a named agent |
amodal/connections/ | API contracts, credentials references, and endpoint policies |
amodal/skills/ | Task-specific procedures |
amodal/knowledge/ | Reference documents and corpora |
amodal/stores/ | Structured data schemas |
amodal/tools/ | Custom actions |
evals/ | Behavior checks |
Source can live in a connected Git repository or in Amodal-hosted snapshots. Changes take effect in production through a deployment.
What Amodal provides
Amodal hosts the runtime and gives your team tools to edit source, deploy agents, inspect sessions and logs, track usage, and run evals. The React SDK connects your application to a deployed agent.
You supply the domain instructions, API access, and permissions. Instructions guide the model; enforced restrictions depend on the tool and authentication configuration. See Guardrails.
Start with the support-agent tutorial or the project layout.