Runtime safety for debugging loops and caps

Guardrails: stop runaway agent runs early.

Guardrails are optional development-time safety rails that abort runs when behavior crosses your limits, then preserve a full evidence trail so root cause is obvious.

Read Guardrails Docs Install AgentDbg

from agentdbg import trace

@trace(
  stop_on_loop=True,
  stop_on_loop_min_repetitions=3,
  max_llm_calls=50,
  max_tool_calls=50,
  max_events=200,
  max_duration_s=60,
)
def run_agent():
  ...

Protect budget

Abort loop-heavy runs before they expand API costs.

Keep evidence intact

Guardrail-triggering events are still recorded for investigation.

Tune per workflow

Set thresholds in decorators, context managers, YAML, or env vars.

Available guardrails

Start minimal, then tighten thresholds as your run profile becomes clear.

Guardrail	Purpose	Typical starter value
`stop_on_loop`	Abort when repeated event patterns are detected	`true` for iterative debugging
`stop_on_loop_min_repetitions`	Minimum repetitions required to trigger loop abort	`3`
`max_llm_calls`	Cap model invocations per run	`25-100`
`max_tool_calls`	Cap tool invocations per run	`25-100`
`max_events`	Cap total timeline events	`100-300`
`max_duration_s`	Cap wall-clock run duration in seconds	`30-120`

What happens when a guardrail fires

The triggering event is recorded.
An ERROR event captures guardrail reason and values.
RUN_END(status=error) is written.
A catchable guardrail exception is raised.

You keep the full timeline to the exact abort point.

Behavior snapshot

Guardrails stopping a run in the AgentDbg viewer

Turn guardrails on for your next debugging run.

Use safe defaults first, then narrow thresholds to match expected behavior.

Install AgentDbg Guardrails Reference