Governance

Policies

Policies decide what an agent may do. Antraft is default-deny: only what you explicitly allow (or a rule permits) can run.

Allow / deny

policy.py
runtime = (
    Antraft.agent(model, tools, task="...")
    .allow(["search_web", "read_file"])
    .deny(["shell_execute"])
    .build()
)
Note
Anything not allowed (and not matched by a rule) is denied. A denied action stops the run.

Decisions

DecisionEffect
allowThe tool runs.
denyThe action is blocked; the run stops.
pauseExecution pauses for human approval (HITL).

DSL rules

Rules add conditional logic on an action's parameters and runtime context.

policy.yaml
allow:
  - wire_transfer
rules:
  - trigger: "action:wire_transfer"
    checks: ["amount > 1000"]
    enforce: pause          # allow | deny | pause
    message: "Large transfer needs approval"
load.py
Antraft.agent(model, tools, task="...").load_policy("policy.yaml").build()

Prerequisites & limits

Policies can require that one action ran before another, and cap totals such asmax_actions. Budgets (tokens and dollars) are covered on the next page.

policy.yaml
prerequisites:
  read_data: ["authenticate"]     # must authenticate before reading
limits:
  hard:
    max_actions: 50

Dynamic updates

Policies can be updated on a running agent (continuous authorization) via the control plane.

shell
PUT /agents/{agent_id}/policy   # new policy JSON
Next →
Budget