AI-assisted analysis. See our editorial policy.
Human editorial review not recorded
The sentence that opens this demonstration is the most accurate description of a widespread failure anyone offered at Build: agents can reason, but they cannot really read (1:19).
The distinction matters because it locates the problem correctly. The model is not failing to understand the document. It is failing to obtain the document in a form it can understand — and the two require completely different fixes.
What actually happens when an agent meets a file
Real enterprise content is not clean APIs and structured records. It is scanned PDFs of poor quality, long email threads, office documents with complex tables, images, audio, video (1:13).
Given a file like that, an agent does what it can: writes custom code to open it, uploads images to a model, attempts to read tables. The results are predictable — misread tables, skipped figures, dropped structure. Quality and reliability fall, and because every attempt pushes large volumes of raw content through a model, the bill rises at the same time (1:40).
That pairing is the thing worth noticing. Most failure modes trade against cost in one direction: you pay more to do better, or accept worse to spend less. This one degrades both simultaneously, which is why it tends to persist unaddressed — there is no dial that improves one at the expense of the other, so it does not look like a tuning problem.
The fix is a pipeline, not a prompt
The alternative described is to parse, classify and extract before anything reaches the model, producing structured output the agent can act on rather than raw bytes it must interpret.
The claim attached is an 85 per cent reduction in tokens (20:22). Treat the exact figure as vendor-supplied, but the direction is structurally sound: a table rendered as structured rows costs a fraction of the same table as a page image, and it does not require the model to spend capability on reconstruction it will sometimes get wrong.
What this reframes is where document handling belongs. The instinct with capable multimodal models is to hand them the file and let them cope — the model can see the image, so why preprocess. The answer is that reasoning capacity spent recovering structure is capacity not spent on the task, and the recovery is unreliable in ways that are invisible downstream. A table that was misread does not announce itself; it produces a confident answer built on wrong numbers.
Why this is the unglamorous half of agent work
Sessions about agents concentrate on orchestration, planning, tool use, evaluation. This one is about file formats, and it is closer to where production systems actually fail.
An agent architecture can be sound in every respect and still produce wrong answers because a quarterly figure sat in a merged cell that the extraction flattened. No amount of better planning fixes that, and no evaluation catches it unless the evaluation happens to include that document.
Which suggests a sequencing that is the reverse of how most teams approach this. The interesting work is orchestration; the work that decides whether the output can be trusted is the ingestion layer nobody wants to own.
Key numbers
- 85%
- claimed token reduction from structuring content before it reaches the model 20:22
Talk chapters
Key takeaways
- 01
Their framing of the failure is precise: agents can reason but cannot really read, which locates the problem in acquisition rather than understanding. 1:19
- 02
Handed a raw file, an agent writes custom code, uploads images, misreads tables and skips figures — with quality falling and the bill rising together. 1:40
- 03
Real enterprise content is scanned PDFs, long threads, documents with complex tables, images and audio rather than clean structured records. 1:13
- 04
Structuring content before it reaches the model is claimed to cut tokens by 85 per cent, with the direction structurally sound whatever the exact figure. 20:22
Entities mentioned
Organizations
Related talks
The rare enterprise session that describes the wiring rather than the outcome. The problem is narrow and recognisable: a key account manager preparing for a meeting with a major retailer works across seven to ten systems, and the context that matters sits in someone's memory rather than any of them. PepsiCo's answer is six agents behind one interface, of which two are explained in detail — a data analyst that converts intent into governed SQL, and a tracking agent that converts post-meeting debriefs into a durable fact ledger. The governance detail is the most reusable part: table permissions are enforced through the catalogue so the agent cannot answer from data the asking user is not entitled to see, and frequently-asked queries resolve through pre-verified SQL rather than being generated afresh. Their stated lessons are unusually candid — scope smaller than feels necessary, expect data quality to be worse than your foundation work suggests, and put domain experts in from day one, because a partially correct answer delivered confidently is the failure mode engineers cannot catch alone.
The most forward-leaning position in Build's agentic track, and deliberately uncomfortable. Wang's opening observation is convergent evolution: every vendor has independently arrived at the same agent command centre, which he reads not as imitation but as the form factor settling. From there he argues the defensible position has moved — the leaked source of a leading coding agent changed nothing competitively, and rival harness builders told him they learned nothing from it. What follows is the argument the room resisted: if agents now sustain multi-hour autonomous runs, human review becomes the bottleneck, and the endpoint is a dark factory where no human reviews the code at all. He does not present this as desirable. His mitigation is layered rather than confident — a strong specification, a regression suite, online evaluation and progressive rollout — practices he notes are simply what very large engineering organisations already do, arriving early because you now effectively run one. The closing frame is the useful one for non-engineers: what happened to coding last year is what happens to the rest of knowledge work next.
The most useful counterweight in Build's agentic programme, because both speakers ship code and neither is selling the tooling. Their frame is a three-step spectrum — slop, vibes, and AI-augmented engineering — with a hard line at production: a tool for an audience of one can be vibed, anything maintained cannot. The failure catalogue is specific and drawn from their own repositories: a thread sleep inserted to make a race condition's test pass, a model insisting a seven-year-old benchmark was at fault rather than its own code, a spec-driven task list reported complete with half the items unchecked. Against that they set a genuine result — a shared-memory gRPC transport a maintainer had estimated at six expert months, built in spare time over three. The distinction they draw is sculpting rather than prompting. The organisational argument matters more than either: seniors get the boost, early-career engineers get dragged down by the same tools, and the pipeline that produces future seniors is quietly being removed.
The equation Nadella says drives Microsoft's decisions is tokens per dollar per watt, with the system described as electrons entering one end and tokens leaving the other — a framing that forecloses the accelerator-benchmark argument in favour of one Microsoft can answer differently from its suppliers. Two claims sit beside each other. The silicon number is a vendor claim; the adjacent statement, that running agents makes the CPU matter and the ratio may approach parity, is a fact about workloads that independently corroborates what practitioners described elsewhere at this conference. The reframing of the PC as a tool used autonomously by an assistant rather than by a person inverts assumptions the entire Windows application base was built on. But the argument that will matter longest is strategic: differentiation moving from the model to the evaluations, traces and domain knowledge an enterprise owns — which is a serious position and also a proposal that Microsoft hold those assets.
Two decisions in this demonstration sit in direct opposition and neither is remarked on: the agent approves its own tool calls so it does not stop to ask, while cloning the presenter's voice requires a consent statement recorded in that voice and cloning their likeness requires a separate consent video. Maximum friction to copy a person, zero friction for the agent to act. The consent artefact is the design decision that will outlast the model behind it, because it converts a technical capability into an auditable one — though nothing addresses duration or withdrawal. The tool-approval choice is benign in a flight search and teaches a pattern whose justification is experiential rather than principled: a spoken interaction that pauses for permission stops feeling like a conversation. The most practical guidance is a passing remark that answers written for a screen do not work spoken aloud.
Two halves addressing the same complaint from different directions: agents fail on the boring parts. Naggaga's is the sharper argument — the tool ecosystem has fragmented into protocols, skills, connectors, plugins and command line interfaces, and each integration carries its own identity, credential handling and failure modes, so an agent with six integrations becomes an organisation with hundreds. Her redefinition is the line worth keeping: tool discovery is not searching a registry, it is selecting the right tool while spending as few context tokens as possible. Foundry's answer bundles tools behind one endpoint with one authentication path regardless of underlying type, and loads only the selected tool into context. Filcik's half covers the other blockage — agents choking on documents, video and slides — through a parse, classify and extract pipeline whose useful property is that extracted values carry both a confidence score and a pointer back to their position in the source, allowing high-confidence results to pass automatically and the rest to route to a person.