Find out what your agents would have been stopped doing.
Run the full pipeline against your own agents with enforcement switched off, and see your real block rate before it can cost you anything. No form, no account, no call. If the answer is that you do not have a problem, that is a useful afternoon either way.
Four steps, in the order you do them
Nothing here is a trial that expires or a key that has to be requested. It is the shipped embedded gateway, configured to observe rather than enforce.
Install
One package. No API key, no account, no network egress — the gateway runs inside your own process on SQLite.
pip install "agentrust-py[embedded,retry]"
Turn it on without enforcing
The full pipeline runs — validation, confidence, risk, decision — and writes every result to the ledger. Nothing is blocked. Your agents behave exactly as they do today.
from agentrust_sdk import embed_gateway, harness
embed_gateway() # in-process gateway on :8765
@harness(agent_id="refund-agent", block_on_block=False)
def refund_agent(user: str, input: str) -> dict:
... # unchangedOr change nothing at all
If you would rather not touch the agent code, the auto-instrumenter patches OpenAI, LangChain and LangGraph in place. Remove it with one call.
from agentrust_sdk import auto_instrument auto_instrument() # roll back with remove_patches()
Read your own numbers
After a representative period, you have the thing you cannot get from a vendor deck: what your agents would actually have been stopped doing, and how often.
agentrust report --since 14d
Four numbers that are yours, not ours
Your real block rate
How many actions would have been stopped, by which rule, and in which agent. Most teams are surprised in one direction or the other, and both are useful.
Risk-tier distribution
How your traffic spreads across low, medium, high and critical — the input to whether enforcement is a small change or a large one.
An audit trail that already exists
An append-only ledger entry with content hashes for every run, from the first day, rather than reconstructed when someone asks.
A latency measurement from your stack
The documented target is under 20 ms on the deterministic path. This tells you what it is on your hardware, with your payloads.
The obvious questions
- Does any of our data reach AgentTrust?
- No. The embedded gateway runs in your own process and persists to local SQLite. There is no API key because there is no account, and no egress to disable because there is none. This is the same component used for air-gapped deployments.
- Will this block anything in production?
- Not in this configuration. block_on_block=False runs the full pipeline and records every decision without acting on it. Enforcement is a separate, deliberate change you make later, if you make it.
- What does it cost?
- Nothing. The core engine is Apache-2.0 and the embedded gateway needs no paid tier. Confidence scoring, risk tiers and central audit are paid capabilities, and the baseline runs without them.
- How do we remove it?
- AGENTRUST_ENABLED=false turns every governance path into a no-op without a code change, or remove_patches() if you used the auto-instrumenter. Uninstall the package if you would rather. A restart is the whole procedure.
- What is the catch?
- The baseline deliberately excludes the parts that need a gateway you do not run on a laptop: the LLM judge, multi-agent trust chain, historical reliability and central analytics. It is enough to tell you whether you have a governance problem, and not enough to run governance at scale.
