brightbeam.works  ·  Papers  ·  Writing  ·  LinkedIn  ·  Work with us

Digital intelligence is becoming abundant, and organisations will soon run thousands of agents alongside their people. Our open-source projects build oversight into the work: every agent has clear limits, a named owner and a record, and nothing it learns spreads until a person approves it. The projects come from the research behind Brightbeam's Continuous Intelligence Operations, the discipline of operating digital intelligence in live work and improving it continuously from the evidence that work produces.

Projects

• CHAP is an open protocol that gives approvals, overrides and handoffs between people and agents one auditable shape, beside MCP and A2A. Overview
• Metis captures what experts notice and do as tacit fragments, each with its source and the conditions for using it. People decide which ones an agent may use. Overview
• Brevet is a local-first Python runtime that turns corrections into proposed rules. A named person approves each one, approved rules ship in signed releases, and any rule can be recalled. Overview

How they connect

A quality reviewer changes an agent's rating from minor to major, because a pump that vibrates during cleaning may have a worn seal. The loop turns her correction into a rule the agent may use, and takes the rule back when the evidence turns against it.

Follow the correction around the loop

1. Sense. CHAP records the agent's rating, the reviewer's change and her reason.
2. Compare. Her ratings are compared with the agent's across similar reports. Four disagree, for the same reason.
3. Surface. Brevet turns the four corrections into one proposal for the quality owner. The agent keeps its current release.
4. Elicit. At a natural pause, Metis asks the reviewer one short question and keeps her answer as a tacit fragment.
5. Decide. The quality owner, not the agent, approves the rule, with tests built from the four reports.
6. Release. The rule reaches the agent in a signed release, and a reviewer still checks its ratings.
7. Monitor. Weeks later, engineers trace the vibration to a faulty sensor, so the rule is recalled and the release flagged.

Anticipate runs ahead of the loop: the earlier reports stay ready to test the model's replacement before the vendor retires it.

Get started

# CHAP: put human and agent decisions on one record
pip install chap-coordinator      # or: npm i @brightbeamai/chap-coordinator

# Metis: capture, review and serve tacit fragments
pip install metis-memory

# Brevet: run the governed learning loop end to end
pip install brevet && brevet demo

Each repository opens with a worked example.

Research

• CHAP · Collaborative Human-Agent Protocol (CHAP), arXiv
• Metis · Tacit Fragments: Operationalising Tacit Knowledge as a Governed Memory Layer for Agentic AI, alphaXiv
• Brevet · Brevet: Change Control for What Self-Evolving AI Agents Learn, ResearchGate

Policy papers, essays and talks live at brightbeam.works.

Work with us

Start with a decision your team knows well. Bring a task where experience changes the result, and the people who do and review it; together we agree how to test where an agent can help. We also work with universities and research partners on proposals and studies.

Discuss your workflow

To contribute, start with the contributing guide in each repository. Questions about the code are welcome in CHAP Discussions.

_{Intelligence, integrated.}