Run one guarded agent call locally first. Add hosted control later when you need retained incidents, alerts, remote kill, and team visibility.
Install with pip install agentguard47. Inspect the SDK on
GitHub.
Fastest first proof
Run one guarded call locally.Install the SDK, set one budget, verify the run locally, then add the hosted control plane later.
pip install agentguard47 openaiFastest first proof
One package. One guarded call. One local proof before you wire any hosted service.
pip install agentguard47 openai
import agentguard
from openai import OpenAI
agentguard.init(
service="repo-coder",
budget_usd=5.00,
trace_file="traces.jsonl",
local_only=True,
)
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "user", "content": "Review this function and return one bug risk."}
],
)
print(response.choices[0].message.content)
Production failure mode
When an agent gets stuck on a broken path, retries pile up, budget climbs, and the team wakes up with no clean incident trail.
A broken tool call keeps firing because the agent never exits the bad path.
Budget keeps climbing while nobody is watching the run.
The team has no retained incident trail when they come back to debug it.
The SDK stops the run locally. The hosted control plane keeps the incident, alerts the team, and gives you remote control afterward.
Stop the bad run. Keep the record. Give the team control.
Put a dollar boundary on a run before a bad agent burns through budget.
Cut repeated tool calls and runaway retry paths before they compound.
Keep the incident, the alert, and the shared record instead of one local file on one machine.
Use the control plane for alerts, retained history, service rollups, and remote kill.
Comparison
AgentGuard is built to stop the bad run first.
Stops bad runs locally, then keeps the incident record and remote controls in the hosted control plane.
Great for tracing and evals. Not the same thing as incident control.
Useful for orchestration and workflow automation. Different job than stopping a runaway production agent.
| Question | AgentGuard | LangSmith | CrewAI | n8n |
|---|---|---|---|---|
| Primary job | Runtime control | Tracing + evals | Framework + AMP | Workflow automation |
| Start local first | Yes | Partial | Partial | Self-hosted workflows |
| Stops a bad run locally | Yes | Not documented | Partial guardrails | Not the primary product |
| Remote kill signal | Yes | Not documented | Not documented | Workflow stop, not agent kill |
| Best fit | Production agent incidents | LLM observability | Agent orchestration | Broad automations |
Based on primary documented product surfaces as of April 9, 2026. We use “Not documented” or “Partial” when the docs do not support a hard “No.”
One package to audit. No runtime dependency sprawl in the SDK path.
The dashboard stores only the traces and events you choose to send.
Local proof first. Hosted later for retained incidents, alerts, and team operations.
Free SDK
Run one guarded workflow first.
Pro control plane
Retained incidents, alerts, and remote kill.
Team control plane
Same control path, sized for a real team.