Agentspecs
Agentspecs are reusable configuration contracts for your agents. They let teams define behavior once, then run the same agent setup across Platform, VS Code, JupyterLab, and CLI.
What an Agentspec Defines
An agentspec can include:
- Agent identity and runtime metadata.
- Model/provider defaults.
- Protocol and transport preferences.
- Tool and MCP server selection.
- Skill selection and runtime options.
- Guardrails, approvals, and sandbox settings.
- Optional parameter schema for launch-time customization.
Why Teams Use Agentspecs
- Portability: move from local experiments to managed runtimes without rewriting configuration.
- Consistency: keep model, tools, and policy settings aligned across interfaces.
- Versionability: track agent behavior changes as explicit config revisions.
- Operational clarity: know exactly which tools, skills, and guardrails are active for a run.
Typical Workflow
- Start from a baseline spec for your use case.
- Add tools, MCP servers, and skills.
- Attach runtime controls (sandbox, codemode, approvals, budgets).
- Optionally define a parameter schema for per-launch values.
- Launch and iterate using evaluation and monitoring feedback.
Best Practices
- Keep a small set of production-grade base specs and derive variants per team.
- Store policy controls in the spec, not only in UI state.
- Use parameter schemas for controlled customization rather than free-form prompts.
- Pair spec changes with evaluations before broad rollout.
Next Steps
- Connect specs to interactive flows in AI Chat.
- Add tool ecosystems in MCP Servers.
- Control behavior with Guardrails.