Guardrails
Guardrails are the limits an agent runs inside. The important word on this page is enforced: what is listed here is what the runtime really does, not a posture the agent promises to keep.
Every agent here already runs metered — each tool call is recorded and each run is charged against a budget cap that stops it rather than warns about it. That is the baseline, and it is what makes an agent safe to leave alone: the worst case is bounded before it starts, not noticed afterwards.
The axis changes behavior mostly in what you can see. A higher-trace posture records more about each step, which costs a little and gives you a run you can reconstruct afterwards instead of a result you have to trust. The trade is straightforward: detail and cost against speed and quiet.
Postures in this deployment
Read from the guardrails registry at page load, with the guarantees each posture actually carries.
| Standard | What every agent here already runs with: metered tool calls and enforced budget caps. The default. Enforced: Every tool call metered to the spend ledger; Per-run/day/month budget caps enforced. | available · default |
| Traced | Standard, plus full trace capture: this agent's governed calls record redacted request and response payloads for later inspection. Enforced: Every tool call metered to the spend ledger; Per-run/day/month budget caps enforced; Redacted request/response payloads captured for every governed call. | available |
How to choose
Standard is the default and it is not a lesser tier — the metering and the budget cap are on either way. Choose traced when you are actively working on an agent and want to be able to answer "why did it do that?" from the record rather than from a re-run.
The honest limitation: these postures bound cost and record behavior. They do not review individual actions before they happen. If you need an agent's every tool call checked against a policy — including agents that don't run here — that is what Agentic Control Plane does, and it is a different product rather than a setting on this one.
Falling back is safe in one direction only, which is the direction it falls: an unavailable posture resolves to standard, which declines to add trace capture and never removes a limit that was already holding.
There is no automated experiment runner yet. Nothing here sweeps this axis for you — changing it means editing the agent and reading the next run. A compare mode that runs one task across several compositions is named on the index as not built, and that is still true.