Guardrails

These guides check what an agent's turn did before its answer is kept, and choose what happens when a check fails.

A guardrail is a rule an agent's turn must satisfy. It's a check and an action. The check is a function over the turn: its answer, its tool calls and their results. The action is what happens when the check fails. Kindgi runs an agent's guardrails once per turn, on its final answer, before the answer is stored.

What a guardrail can do in 0.1

• Only halt acts. A failed halt guardrail stops the turn with guardrail-violation. Any other action is recorded on the turn as a violation, and the turn completes: retry, escalate and compensate don't run the turn again, escalate or call a tool. So a guardrail can't block an answer and have the agent try again. With halt, the person or your app sends the message again.
• A check sees one turn: the turn's user message, its tool calls and their results, and its answer. Earlier turns aren't in it, so a check can't compare an answer with what an earlier turn said or fetched.
• What a check keeps in memory doesn't last. It lives in the process that runs the pack's code. kindgi dev loses it on every reload and restart, and a deployment loses it whenever an instance stops. With more than one instance, each keeps its own.

For a check on figures, tell the agent in its instructions to call its tools again in every answer that quotes a figure: the figure is then in that turn's tool results. That depends on the model following the instruction, and it doesn't cover figures the user typed in an earlier turn.

The guides

• Write a guardrail: the check, the declaration, and wiring it onto an agent.
• Configure a guardrail: settings the check runs with, and their schema.
• Stop a turn or record a violation: halt versus log-only, severity, which agents a guardrail covers, and how a guardrail is doing: how often it passed, was violated, blocked a turn or couldn't run.
• Use a built-in check: name one of Kindgi's seven checks instead of writing your own, and the settings each reads.

The examples use the quickstarts' pack, my-pack, whose agent answers with dev-echo until you connect a model.