# `Parlora.Backends.Sagents`
[🔗](https://github.com/hfiguera/parlora/blob/v0.1.0/lib/parlora/backends/sagents.ex#L1)

Runs one isolated, bounded Sagents execution per admitted request.

Install optional `sagents == 0.17.1`. Configure `:agent` as a function receiving
a BackendRequest and returning `{:ok, agent, state}`, `:result` as a function
mapping a successful terminal tool result to `%{status: ..., facts: ..., summary: ...}`,
and `:completion_tool` as its name. `:max_runs` defaults to 8 (maximum 100).

The application constructs the model, credentials and synchronous tools. Build
the agent with `replace_default_middleware: true` and empty middleware; default
filesystem/subagent tools, custom modes, fallbacks and asynchronous tools are
deliberately unsupported. This makes execution ownership and authority explicit.

Cancellation is advisory: it stops local execution and reports an uncertain
external outcome. It never confirms reversal of an already executed tool.
Updates, model token streams, progress and clarification are not advertised.

---

*Consult [api-reference.md](api-reference.md) for complete listing*