Behavry

---

Behavry is the system of record for autonomous agents. It sits between an agent and every system it touches, binds each action to an identity, and reconstructs what happened into a signed Decision Trace. Because it is in the path, it can also permit, redact, hold or block an action against policy.

This repository shows what that looks like in real workflows. Every example is a complete, small agent that does a useful job, with the decisions that make it trustworthy written down next to the code.

Examples

Example	What the agent does	Systems	Guide	Code
Meraki Alert Triage	Triages network alerts, writes a structured note to the PSA ticket, and asks a technician before it touches a device	Cisco Meraki, Halo PSA, an MCP gateway, Claude	PDF	Python
Microsoft 365 User Lifecycle	Handles new-starter and leaver requests: locks leavers out at once, and holds every account creation and mailbox change for a technician	Microsoft 365, Halo PSA, an MCP gateway, Claude	PDF	Python

What every example includes

examples/<name>/
  README.md      one-page overview: what it shows, systems, Behavry features used
  guide.md       the guide, written for the team that will run it
  guide.pdf      the same guide as a branded PDF, ready to share
  diagram.svg    how the agent's calls flow through Behavry
  screenshots/   where each step is configured in the Behavry dashboard
  agent/         the runnable code, with its own README, tests and Dockerfile

policies/        the Behavry policies the examples run under, with OPA tests
tools/           the PDF builder and the screenshot capture script

Each guide follows the same arc: what the example proves, how to scope a first trial, how the agent is built and why, the policy it runs under, and how to tell whether it worked.

The patterns they share

The examples differ in what they automate. They share a few rules that make an agent's work something a client, an auditor or your own team can rely on:

• Every call goes through Behavry. Model calls go through Behavry's model proxy and tool calls through its MCP gateway, so the record has no gaps.
• One label per run. Each piece of work (an alert, a request, a ticket) carries a run label on every call, so it reads as one Decision Trace.
• The agent has its own identity, in Behavry and in every system it uses, so its actions are attributable and it can be paused on its own.
• Untrusted input stays data. Text from outside is fenced in the prompt, the model returns fixed fields, and the agent's code, not the model, decides which tools to call.
• Actions wait for a person. Anything that changes a system the agent does not own is held by policy until someone approves it, and the outcome is recorded either way.

Run an example

Each example's agent/README.md has its setup. In general you need:

1. A Behavry plane and an agent identity (client_id and client_secret).
2. The target systems registered in Behavry, with a policy for the agent.
3. The credentials the example names, kept in your secret store.

Then:

cd examples/<name>/agent
pip install -r requirements-dev.txt
pytest

Every example's tests run without any live system, so you can read and change the code before connecting anything.

Build the PDFs

The PDFs are generated from each guide.md, so the guide and the PDF never drift apart.

pip install -r tools/pdf/requirements.txt
python -m playwright install chromium   # skip if Google Chrome is installed
python tools/pdf/build.py examples/*/guide.md

Contributing

New examples and improvements are welcome. See CONTRIBUTING.md for the layout an example needs and the checks it must pass.

Security

Please report a security issue privately to [email protected] rather than in a public issue. These examples are references: review them, and run them against test systems first, before you point them at anything that matters.

License

Apache 2.0. The Behavry name and logo are trademarks of Behavry and are not covered by the license.