How I read the field without believing it

Field readings, one of a series. Before the series reports what the field says, this piece says how the reading is done, what it costs, and where it went wrong once in a way worth keeping.

The pile

Since July 2026 a research intake has run beside my governed agent-platform prototype. Six streams feed it: Medium, Substack and arXiv, scanned weekly against a list of twelve topics; the Cloud Security Alliance’s papers; Dutch institutional primaries such as VNG frameworks and ministry letters; and the carousels and posters that circulate on Instagram and LinkedIn, which I add by hand when one is doing the rounds. Today it holds about nine hundred sources, 871 of them with a verbatim transcript, and sixteen topic briefs that summarise what moved, refreshed every week.

The size is not the point. Anyone can collect Medium. The point is three reading rules, and the register that enforces them.

Rule one: open the artefact

On 11 July 2026, every image in the collection was transcribed for the first time: 91 images across 27 folders. Until that day no agent in the pipeline had opened a single one. The triage had rated eight sources as image-only, unreadable, or without accessible content, and marked two of them for discard. All of it was wrong. The transcripts came back with zero illegible markers.

The cost was specific. Two of the eight “unreadable” sources turned out to be the origin of two chapters of my own platform documentation: the ten-component harness anatomy, and the five-layer governance model. Both had been drawn from the images, one to one, in order, by a human who had looked. The pipeline that was supposed to check them had concluded the folders were empty because the file beside the image said nothing.

And the captions were worse than useless. Every owner caption checked against its own image was wrong about it. One called a row a column. One dropped an entire band from a chart. One named an industry that was not on the chart. One invoked four regulations the image never mentions. One claimed three capabilities the slides never state.

The rule that came out of it binds every document in the register: before recording that a source has no content, open it. A file format is not a finding. An absence recorded without a search is an assumption wearing a finding’s clothes. Wittgenstein, in a different argument, put the whole method in four words: don’t think, but look.

Rule two: a figure carries its provenance

A specific percentage attached to a famous firm’s name is the most propagable object in the whole corpus. It survives being copied into a slide long after its provenance is gone. So every number that arrives is recorded with exactly what its source attributes it to, at the moment it becomes visible rather than after it has spread. Three dispositions, and the difference between them is the useful part.

Traced: someone located the primary, read it, and recorded what it actually says. A checkable lead: a real primary exists and nobody has opened it; the right response is to open it. Unattributed: no method, no sample, no locatable source. It stays the source’s figure, reported as what the source states, never restated as mine.

The method has already corrected a figure three times. A practitioner guide to multi-agent orchestration cited a real NeurIPS paper on why multi-agent systems fail, and reported its failures as split evenly, a third each, across three categories. The paper exists, with 1,642 annotated traces across seven frameworks. The even split appears nowhere in it: the per-mode percentages sum to roughly 44, 32 and 24, and the category names in the guide were the author’s paraphrase. An engineering post from a model vendor says multi-agent systems use about fifteen times the tokens of a chat; it also says single agents use about four times, which makes the multi-agent premium over a single agent closer to four than fifteen, and the same post’s observation that token use explains most of the variance is the vendor’s argument for the architecture, not a concession against it. And an essay’s “earlier Anthropic study” turned out to be two separate reports, four months and two teams apart, with one of its four figures describing a different quantity than the essay claimed.

None of that makes the guides worthless. The orchestration guide’s failure modes are good, and I use them. The deck that carried “95% of executives experienced at least one AI-related incident in production” has a five-layer governance model I think is right. The deck may be right about the architecture and still be unable to say where its number came from. I take the architecture and I say, of the number, exactly what the deck said: nothing.

Rule three: nothing is promoted because a source says so

A source may corroborate a decision that already exists. It may raise a question. It may never make a decision. Sixteen digest passes, eighty sources fully digested, zero promotions: not one architecture decision in my platform was taken because a paper or a post said so, and each decision names the sources that corroborate it and the ones that collide with it. Collisions are written down, not resolved by seniority. A beginner’s primer on knowledge graphs contradicts two of my own corrections; the register says so, and the primer stays on the shelf.

The rule has a second edge. Three sources that agree are not three votes if they share an ancestor. The vendor post above, a Stanford preprint that critiques it, and a later paper built from the vendor’s own published guidelines are one lineage, not three independent confirmations. The register records that too, because the moment you forget it is the moment you write “the field agrees”.

What it cost, and what it buys

The honest accounting includes the ceremony. For its first weeks the apparatus downgraded sources on principle: “practitioner post”, “vendor blog”, “none is citable”. That felt like rigour. It was mostly ceremony, and it did measurable harm, because it was the same apparatus that declared the eight sources unreadable without opening them. The rule that replaced it is not a weaker gate but a different discipline: read the source, record what it says, say where it is wrong, go deeper where it is thin.

What it buys is a brief per topic that a team can read before spending a research pass from scratch, a register that has caught itself and is therefore worth more than one that has never been tested, and this series. Each field reading will say what changed in one topic, from three to five sources, where a source is wrong or thin, and what it means for a CTO or a municipality. I will link every original, name every author, and redraw nobody’s chart.

Have you tried opening it? It is the IT Crowd’s question with the file in place of the router, and it has fixed more of my register than any policy did.

References

The allusions and quotations in this piece, with their origin:

• The IT Crowd (Channel 4, 2006-2013): “Have you tried turning it off and on again?”
• Ludwig Wittgenstein, Philosophical Investigations (1953), §66: “don’t think, but look!”
• Mert Cemri, Melissa Z. Pan, Shuyi Yang et al., “Why Do Multi-Agent LLM Systems Fail?”, arXiv:2503.13657. The category sums are my additions of the paper’s per-mode percentages.
• Anthropic, “How we built our multi-agent research system”, 13 June 2025: “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.”
• Gartner press release, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, 25 June 2025. The one firm-and-year attribution in the corpus that traces to a locatable primary.