Production AI agents
your regulator can sign off.
Runtime governance for production AI agents. CPS 230 evidence on Day 1, EU AI Act and ISO 42001 mappings included. Built for APRA-regulated financial services in Australia and APAC.
The governance layer your CISO already asked you to build.
Three things every APRA-regulated team needs the day they put a non-trivial AI agent into production.
Built for APRA CPS 230 + EU AI Act
Controls mapped directly to CPS 230 operational risk, EU AI Act high-risk obligations, and ISO 42001. Every event in the log carries the framework reference. Your auditor opens one PDF and stops asking questions.
Runtime enforcement, not prompt-scanning
AgentGuard sits in the agent runtime — it intercepts tool calls, model outputs, and dollar-spend before they happen. Policies are YAML, not regex on prompts. Block actions, require human approval, or escalate to a second agent.
Audit evidence your board accepts
Every decision your agents make produces a signed, Bitcoin-anchored audit record. The compliance pack assembles them into a single PDF with framework mappings, control owners and statistical sampling — the format your Big-Four auditor expects.
Four steps from pip install to audit-ready.
Install
One line in Python or Node. Works locally and behind your firewall — no SaaS dependency for runtime.
pip install agentguard-tech
Wire your agent
One decorator or middleware per framework. We ship adapters for OpenClaw, LangChain, CrewAI, AutoGen, OpenAI Assistants, and MCP.
from agentguard import guard @guard(policy="cps230") def run_agent(...): ...
Author policy YAML
Declarative policies covering tool whitelists, dollar caps, PII egress, human-in-the-loop gates, and after-hours rules.
limits:
daily_spend_aud: 500
tools_allow:
- search
- read_file
require_approval:
- send_emailGenerate evidence
One command builds the signed compliance pack — the same PDF format APRA-regulated boards already accept.
agentguard evidence build --framework cps230
Govern OpenClaw fleets natively.
OpenClaw is the most popular open agent framework in the world. Most of them run unmonitored. AgentGuard is the only commercial governance product that natively manages OpenClaw — written by an operator who runs an 18-agent OpenClaw fleet in production.
- Drop-in plugin — no fork, no patches.
- Per-claw policy: dollar caps, tool whitelists, human approvals.
- Fleet view: which claws are live, what they're doing, what they cost.
- Killswitch and rollback on policy breach.
# Drop into your existing OpenClaw config
plugins:
- name: agentguard
package: "@the-bot-club/agentguard"
options:
policy: ./policies/cps230.yaml
evidence: ./evidence/
anchor: true # Bitcoin-anchor audit roots
fleet_id: bnb-prod # tag for fleet view
tools:
allow: [search, read_file, http_get]
require_approval: [send_email, transfer]
limits:
daily_spend_aud: 500
max_parallel_claws: 12One event spec. One policy DSL. One audit log.
AgentGuard governs your production agents. Tribunal governs your coding agents. Together they're one spec across every agent your company runs — from the CrewAI worker handling claims, to the Claude Code session shipping the patch that powers it.
Free to start. Flat-fee to ship compliant.
Free
- 1 agent, 7-day retention
- Local audit log
- Community Discord
- MIT-licensed SDK
Team
- Up to 25 agents
- 90-day retention
- SSO (Google, Microsoft)
- Signed audit roots
- Email support
Compliance
- Unlimited agents
- 1-year retention + cold archive
- CPS 230 evidence pack
- EU AI Act + ISO 42001 mappings
- Bitcoin-anchored audit roots
- Quarterly readiness review
Enterprise — on-prem, SLA, dedicated SE. Talk to us →