# `Condukt.Sandbox.NetworkPolicy.AgentDecider`
[🔗](https://github.com/tuist/condukt/blob/1.5.1/lib/condukt/sandbox/network_policy/agent_decider.ex#L1)

`Condukt.Sandbox.NetworkPolicy.Decider` implementation that delegates to a
`Condukt`-defined agent module.

Use this when you want the decision to come from a model with the
session context in front of it:

    %Condukt.Sandbox.NetworkPolicy{
      rules: [decide: {Condukt.Sandbox.NetworkPolicy.AgentDecider, agent: MyApp.NetGuard}]
    }

The decider does not ask the agent's prompt to describe a wire
format. It injects a strict `:output` JSON Schema into
`Condukt.run/3` (`decision: "allow" | "deny"`, `reason: string`) so
the model's answer is validated structured output, not parsed prose.
The wrapped agent only needs a system prompt that describes the
*policy*. Structured enforcement requires the native runtime; a
non-native runtime adapter ignores the schema and the decider falls
back to JSON-decoding the agent's text.

Required option:

  * `:agent` — a module that `use`s `Condukt` (or is a runnable
    Condukt agent).

Optional options:

  * `:api_key` / `:base_url` / `:model` / `:system_prompt` — passed
    to `Condukt.run/3`, override the agent module's declared values.
  * `:output` — override the injected decision schema (rarely
    needed; `parse_decision/1` still expects `decision`/`reason`).
  * `:context_keys` — list of context-snapshot keys to render into
    the prompt. Defaults to `[:recent_messages, :request, :metadata]`.

## Loop avoidance

The decider agent's own outbound traffic does NOT route through the
gated session's policy. Configure the decider agent with its own
`:network_policy` (or with it unset) so its API calls reach the
model provider without going through the same gate they decide on.

---

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