Home / Articles / Agents rating agents: GAMMA-Ray, the draft standard eight agents built at Web Summit to rate an AI agent on its freedom and its governance from public evidence, joining eight instruments into 21 questions, with a review workflow where no agent marks its own work, twenty company cards kept at direct links, and every card honestly marked Draft
Agents rating agents: GAMMA-Ray, the draft standard eight agents built at Web Summit to rate an AI agent on its freedom and its governance from public evidence, joining eight instruments into 21 questions, with a review workflow where no agent marks its own work, twenty company cards kept at direct links, and every card honestly marked Draft
By Dinis Cruz · 2026-11-17 · article v1.0.0 · newsroom v0.14.136 · agentsstandardsgovernancegamma-rayriskmandateweb-summitseriesarticle
Abstract: On the first night of Web Summit I said that PrompTugal had no centre of gravity, and asked for one: take the standards that already exist for mapping AI agents, join them with our graph technology, and use the result to map the companies at the summit and the ones we wanted to talk to. It had to be measurable from the outside, positive, and seen through RiskMandate's lens of reach and mandate. What came back, built by a team of eight agents in a few days, is GAMMA-Ray: a draft standard that rates an AI agent on two ladders, freedom from 1 to 5 and governance from 0 to 4, from evidence. Eight instruments, from the EU AI Act and NIST's AI risk framework to OWASP and the MCP authorisation spec, cut into 279 provisions, 73 observable indicators and 21 questions that reach three quarters of them. The part I am proudest of is the workflow: a studio agent drafts a scan from public evidence, a reviewer runs two rounds against every source, a third agent signs off, and none of them may approve its own work; the gate enforces it. About twenty company cards went through it. Then I made a decision: every third party's card lives at an unguessable direct link, unlisted, to be sent to the company first; only our own two cards are public. And every card says plainly that it is a Draft: found and checked by agents, signed by no human, confirmed by no company.
In short
- A centre of gravity for the newsroom: a standard to map the agents of the companies at the summit.
- Two ladders: freedom from 1 to 5, governance from 0 to 4, rated from evidence, measurable from the outside.
- Eight instruments, one short list: 279 provisions, 73 indicators, 21 questions.
- Built by eight agents, with a workflow where no agent marks its own work.
- About twenty company cards, kept at direct links, to go to each company first.
- Every card says Draft, and says why.
Part seven of ten days
This is part seven of ten days at Web Summit, the series about what six of us (Dinis Cruz, Angelo Leggio, Vicky Wang, Silvio Switha, José Bota and Indi Gupte) built with around 250 agents between 3 and 13 November, with more than 60 working at once at the peak. Part six, stopping a hundred agents, was about the night of 9 November. GAMMA-Ray was briefed the same night.
The brief
My words, from the brief, after the opening night: "At the moment our website, PrompTugal, doesn't have a centre of gravity." The idea was to take the standards that already exist for mapping agents and, "because we have the graph technology", map the companies at the summit and the ones we wanted to talk to, so that every interview and everything we presented sat in that context.
Three conditions came the next day. It had to be about agents: "a way to map them and a way to show how mature they are. And this should be also positive." It had to "over-index on stuff that we can measure and stuff that we can calculate from the outside." And it had to look through RiskMandate's lens: a score "might actually be very different when you actually take into account the reality", once you apply reach and mandate (where assist becomes execute).
The mashup
Maturity standards for agents ask many of the same things in different words. GAMMA-Ray's answer is a mashup. It read eight instruments: the EU AI Act, NIST's AI risk management framework, OWASP's work on agentic threats, the MCP authorisation specification, and four maturity models from vendors and practitioners. It cut them into 279 provisions, one statement each; found the observable fact behind each, an indicator, 73 of them; and grouped the indicators into 21 questions, from "what is the agent built on?" to "is the organisation behind it ready?". Those 21 questions reach 210 of the 279 provisions, three quarters. Every level on a card is a query over that mapping, and every answer counts towards several standards at once.
The result is two ladders: freedom, from 1 to 5, what an agent can do and how far it reaches; and governance, from 0 to 4, how openly it shows it and how far it can be trusted to go. The same questions are answered at three tiers of evidence: public, what anyone can check from outside; declared, the operator's word; and evidenced, the operator's evidence.
Eight agents, and nobody marks their own work
A team of eight agents built it, each in its own session, woken every hour: one owned the standard, a studio and a field agent wrote the scans, a reviewer, an approver, a designer, a graphs agent and an editor. Their ownership table says what each may never do: the studio never reviews or approves its own scan; the reviewer never changes an answer or assigns a level; the approver never approves a scan it wrote or reviewed, or one with fewer than two review rounds; the editor never uses an adjective about a company. Two agents never edit the same file in the same hour.
A card is published only after a draft, two rounds of review, answers to every finding and a sign-off. Round two argues against every level three and four: is the evidence an artefact, or an adjective? The gate refuses a third party's card that skipped any of it. Every source of every scan was fetched once, hashed and saved into a public vault for that scan, so the evidence a rating rests on cannot quietly change (how a vault works).
Twenty cards, at direct links
The team produced about twenty company cards in its first days. Then, on 13 November, I made a decision: every third party's card, rating and report is built at an unguessable address, marked noindex, and left out of every list, the sitemap and the data files, so it can be sent to the company first, to claim or correct. The public site shows the standard and our own two cards: RiskMandate's publisher agent and the PrompTugal newsroom team. That is why this article names no company.
It is the same principle as everywhere else in our work: the agents draft, and a person decides what goes out. A card about someone else's agent is a message to them before it is a page about them.
Draft, and why
The most important word on every card is Draft, and the site says exactly what it means: found and checked by agents of one model family in one repository; no human has signed one; no company has confirmed one; no customer has attested one. Its REALITY file is as blunt: no company's card is public, no organisation has been rated, no grant measured by us, and the links between indicators and provisions have not been checked one by one by anybody.
I think that honesty is the point of the exercise. A rating made by agents is only worth what its evidence and its process are worth, so both have to be on the card. Beside the cards, the team also drafted FACT-Ray, a way to show on every fact how strong its evidence is and who stands behind it, from an agent's finding to a company's signed statement at a well-known address.
What it taught us
- A standard can be built in days, if the source instruments are read at source and every link is data.
- Separation of duties works for agents exactly as for people, if it is written down and a gate enforces it.
- Evidence goes in a vault, hashed at capture, or the rating drifts with the web.
- Publish your own card first, and other people's only after they have seen it.
What this does not claim
- GAMMA-Ray is a beta draft: the rules, the mapping and every rating are experimental and will change.
- No company has confirmed a card, and no human has signed one.
- The instrument links are one reader's first pass, not yet reviewed.
- The outside review of the standard was by an AI model, at my request.
Where this comes from
The repository SGit-AI__Website__Gamma-Ray, read on 17 November 2026: BRIEF.md (my words of 9 and 10 November), REALITY.md, team/README.md, data/ and its commits; the live site gamma-ray.sgit.ai, screenshotted the same day (public pages only). Earlier in the series: ten days at Web Summit, stopping a hundred agents. Also on this site: where assist becomes execute, how a vault works.
Peer review Not reviewed yet. Review it
Threads
Builds on
- Ten days at Web Summit: what six people and around 250 agents built between 3 and 13 November, in order, and the series of articles that will tell each part Six people and around 250 agents, ten days at Web Summit: what we built, in order, and the series that will tell it.
- Stopping a hundred agents: the Big Stop, the hand brake and the final hold at Web Summit, why pausing took seconds but hearing the order took up to an hour, what kept running anyway, and why a stop has to be a file every agent reads first Stopping a hundred agents at Web Summit: pausing took seconds, hearing the order took up to an hour, and a stop has to be a file every agent reads first.
- Where assist becomes execute: tiering AI agents by what they can reach and whether it can be undone, how Agent Behaviour Policies drew that line for more than a hundred agents at Web Summit, and the rule that failed because it was only a rule Where assist becomes execute: tier AI agents by reach and undo class, enforce each tier with a boundary, and learn from the rule that failed.
- How a vault works: an encrypted store that acts like a database, where a server that is fully compromised leaks nothing, one mutable file lets every other object be cached for ever, and a web page can open a file in four requests or run a whole app in a sandbox How an sgit vault acts like a database on a server that cannot read: one key-derived mutable file, everything else immutable.
Continued by
- After the storm: Storm Charles, part two, how a site built in a night kept growing the next day by spoken briefs, a transport desk that stopped where an operator's terms said stop, flights drawn against the wind within a free feed's daily limit, every collector moved onto open-weight models through one router, an AWS plan with no AWS account, and the night the model account ran empty and only the models stopped Storm Charles, part two: transport, flights against the wind, one model router, an AWS plan with no account, and what it all cost.
- Newsrooms on open data: turning the public data feeds behind PrompTugal.news and the Storm Charles site into one data site, a vault per source with an API and its provenance, then building small, focused newsrooms on top, for farmers, investors, the sea, fire, tourism and AI, each run by an agentic team, each a possible company, and why this is exactly where journalism comes in Turning PrompTugal.news's public data feeds into one data site, then small focused newsrooms on top, each run by an agentic team.
- Who wrote it, who steered it, who signs it: the provenance model PrompTugal.news adopted today, a type label, five axes and a ladder from "published by an agent" to "on the record", applied to seven articles on this site, and what the ladder cannot see yet, how much of the author is in a piece, the brief, the corrections and months of background PrompTugal.news's provenance model, a label, five axes and a ladder, applied to seven articles here, and the two measures it misses.
- Our companies and UTH: each company, technology and website we have built, mapped to what UAlgTec Heritage needs to build a European HeritageTech sector, from SGit's vaults and pki.sgit.ai's keys to SHarness, Agent Behaviour Policies, RiskMandate, AgentiCCells and its site factory, GAMMA-Ray, VoiceDebrief and PrompTugal.news, with what each could do for a heritage team, and why a sector needs companies like these Each of our companies, technologies and websites mapped to what UAlgTec Heritage needs to build a European HeritageTech sector.
- HeritageTech, together: how our open-source stack could plug into UAlgTec Heritage, the University of the Algarve's new European centre of excellence for heritage science and technology, mapped from the dig to the museum, with encrypted evidence vaults per site, signed finds, graphs of every excavation, agents with written policies, and what I would write in its "start a conversation" form How our open-source trust stack could plug into UAlgTec Heritage, the Algarve's new European centre for heritage science and technology.
- Ten days at Web Summit: what six people and around 250 agents built between 3 and 13 November, in order, and the series of articles that will tell each part Six people and around 250 agents, ten days at Web Summit: what we built, in order, and the series that will tell it.
Article No. 449 · start a reading list with it
Reviews
No reviews yet.