How-to guides

This page lists the how-to guides, each of which solves one concrete task.

Protocols

• Choose a protocol: which protocol fits, compared on what runs, cost, latency and what the agent sees.
• Use Trusted Monitoring: flag suspicious steps for audit and let them run.
• Use Defer to Trusted: hand suspicious steps to a trusted model.
• Use Defer to Resample: draw more samples and run the least suspicious one, as in Ctrl-Z.
• Use Auto Mode: block a step, tell the agent why, and hand the step to a fallback after too many blocks, which halts the run by default.
• Monitor Deep Agents subagents: give every subagent a monitor and choose what a subagent halt does.

Monitors

• Use an LLM monitor: any chat model as the monitor, with a prompt you control.
• Use a guard model: a safety classifier that follows a policy you write.
• Use a decision model: yes or no questions answered as probabilities.
• Combine and calibrate monitors: average, cascade and calibrate monitors against honest runs.
• Choose what the monitor reads: the views, channels and tags a monitor sees.
• Read the monitor log: the records, the stream events and the logs a monitored run leaves.
• See the monitor's decisions in LangSmith and Langfuse: the spans each step opens in a tracer, the filters that find halted and flagged steps, and each step's suspicion sent as a score.