The research behind Brightbeam’s Continuous Intelligence Operations, designed for a world where digital intelligence is abundant.
Who we are designing for
Digital intelligence is becoming abundant. Each better model moves more work to agents, and organisations will soon run thousands of them alongside their people.
Governance has to scale with them.
Thousands of agents can’t be checked one by one, so oversight has to be built into the work. Every agent has clear limits, a named owner and a record, and nothing it learns spreads until a person approves it.
We design for a world where people and agents work in harmony, as a human‑agent symphony.
01The problem
Experts notice what matters, spot the exceptions and adjust when things change. Very little of that reaches a system.
Most AI works in the compliant core, automating what the systems already know. The judgement that decides the hard cases sits in the other three.
Procedures, batch records and the systems that hold them.
Handovers and sign-offs whose reasoning leaves no trace.
The spreadsheets, extracts and notes people keep for themselves.
What experienced people notice and do, and never write down.
What happens to everything your experts know when thousands of agents arrive who never met them?
02The loop
Take a manufacturing site where an AI agent sorts equipment reports and a quality reviewer checks its ratings. A pump vibrates during cleaning. The agent rates the report minor, and the reviewer changes it to major because she suspects seal wear. Follow that one correction through the loop, with CHAP, Metis and Brevet each playing a part.
Each stage adds a line to the record. Lines signed by a person are marked in ember.
CHAP records the agent’s rating, the reviewer’s change and her reason.
Her ratings are compared with the agent’s across similar reports.
Brevet turns the four corrections into one proposal for the quality owner.
At a natural pause, Metis asks the reviewer one short question and keeps her answer as a tacit fragment.
The quality owner, not the agent, approves the rule.
The rule reaches the agent in a signed release, and a reviewer still checks its ratings.
Weeks later, engineers trace the vibration to a faulty sensor, so the rule is recalled.
The earlier reports are kept ready to test the model’s replacement before the vendor retires it.
CHAP keeps the decision record throughout: what the agent proposed, who changed it and why. Try this example on the Brevet page
03The discipline
CIOps is the discipline of operating digital intelligence in live work, and improving it continuously from the evidence that work produces.
It connects people, agents and systems so that decisions are made well.
Toyota builds quality into the line. Anyone can stop the line when something is wrong.
DevOps builds, tests, releases and watches software continuously, and the team that builds it also runs it.
CIOps captures, tests, approves, releases and watches judgement continuously, with people and agents together.
It never stops, because three things keep moving
A regulation is published. A supplier changes a material. A sensor drifts out of calibration.
Experienced people leave and their successors decide differently. Reviewers stop checking an agent that is usually right.
A vendor retires a model and its replacement behaves differently on the same inputs. Some agents also change themselves between upgrades, by keeping what worked. They learn, and nobody signs.
Governable decisions
Every decision is auditable, traceable, defensible and explainable.
More importantly, the way the loop runs means the work produces the signals that improve it. Decisions, evaluations and human judgements are captured as a by-product of running Continuous Intelligence Operations.
We help each organisation run it, so its decisions stay sound as the world, its people and its models change.
Regulated industry
Regulated operations already run on records and signatures. A deviation is logged, every change has a named owner and a person signs the release.
The EU AI Act now asks the same of high-risk AI, with automatic records, monitoring after deployment and people who can intervene. We design to those duties and to GxP practice.
A year on, an inspector asks why the pump rating changed twice. The reviewer has moved site and the model has been replaced, but the record still shows what was known, who decided and why.
The record outlasts
the people and the models.
04The research
CHAP
CHAP gives approvals, corrections and handovers a common structure, so a reviewer can trace a decision across the people and agents involved. The reference servers and conformance tests are open source.
Metis
Metis keeps a tacit fragment: an account of what someone noticed and how they responded, with its source and the conditions for using it. A separate review decides whether an agent may use it as guidance.
Brevet
Approved rules reach the agent in signed releases that also record its set-up, so a team can trace a change and withdraw it if the evidence turns against it.
05Research agenda
Work with us
Show us a task where experience makes a difference. We can work through it with the people who do and review the work, identify where an agent might help and agree how to test the result. We also work with universities and other research partners on proposals and studies.
