Your AI Agents Can Query Everything Now. Except What Nobody Ever Wrote Down.

Eleven seconds. That is all a newly connected agent needs to answer a question that used to cost a week of asking around, and the answer arrives formatted, sourced and traceable, with every retrieval call written to an audit log a regulator would be glad to read. Ask it why the settlement job skips accounts flagged D7 on the last business day of the quarter. It will tell you. It will also be wrong, because the reason was never a rule anybody recorded: it was a workaround for a downstream partner's reconciliation defect, explained once, in a corridor, by an engineer who retired eight months ago.

None of that is the context layer's fault. It did exactly what it promised: fetch what exists, respect permissions, log the call. Not a bug. The gap it fell into was never a plumbing gap at all, but the distance between what your systems hold and what your organization actually knows, and 2026 has been a very good year for closing the first one while leaving the second sitting exactly where it always sat.

What did 2026's wave of AI agent context layers actually solve?

A real problem, and it solved it properly. Before governed context layers, wiring an artificial intelligence agent to requirements data meant one of three bad options: paste the ticket into a prompt by hand, scrape the repository with a service account nobody governed, or let the model guess. All three leak. The first is slow and lossy, the second is a security incident waiting for a date, and the third produced the specific failure that made 2025 exhausting, where an agent invented an acceptance criterion no human had written and then built against it with total conviction.

So the tooling vendors answered, and they did it well. The Model Context Protocol, an open standard for handing an agent structured, permissioned access to a system, became the shared plumbing. Jama Software walked through Jama Connect MCP publicly on June 30, 2026: an engineer can pull approved requirements and their governance status straight into the development environment, and the company is explicit that a change made this way is still permission controlled, version controlled and recorded in audit logs. Visure Solutions had shipped the VISURE MCP Server a few weeks earlier, on June 4, giving agents structured reach into requirements, risks, verification evidence and compliance relationships under comparable controls. Both are useful. Any team refusing this category on principle will simply be outrun by the teams that adopted it.

Now notice what the vendors claim, and what they carefully do not. Read the wording. Jama frames the value as access, traceability and audit readiness and stops short of promising that any given requirement is complete; Visure describes secure interaction with engineering context rather than a judgement about that context. Neither company is overselling here. Their market is.

Why can a governed context layer still be confidently wrong?

Because retrieval has no concept of absence. Ask a well-built context layer a question whose answer was never recorded, and it does not hand back an empty set with a note explaining that this organization never wrote the thing down. It returns the closest match it found, ranked by similarity, and the model on the other end turns that into paragraphs. Fluency is not evidence. The failure is silent by construction, which is precisely the property that makes it expensive.

Two separate 2026 evaluations put numbers on this, and neither flatters the field. RefusalBench, published at the European Chapter of the Association for Computational Linguistics conference in March 2026, generates diagnostic test cases by perturbing context in 176 distinct ways across six categories of informational uncertainty, then runs more than thirty models through them. Refusal accuracy fell below 50% on multi-document tasks. Below half. The authors describe frontier models as exhibiting dangerous over-confidence or over-caution, and report that neither scale nor extended reasoning improved the result.

The second study asks the agentic version of the same question. In AgentAbstain, submitted on July 11, 2026 by a team at the University of Illinois Urbana-Champaign, seventeen frontier models were run through 263 paired tasks across 42 executable sandbox environments, where each pair holds one task the agent should perform and a near-identical variant it should refuse. The best of them scored 59.5% across both halves. Not one cleared 60. Read that paper's second finding twice, because it is the one that matters: abstention capability turns out to be largely independent of general task-solving ability, which means a model that gets better at the work does not thereby get better at knowing when the work cannot be done.

Those two numbers are the whole argument compressed. Your context layer got faster, cleaner and considerably better governed this year, and the thing consuming it did not get proportionally better at saying the three words that would actually protect you. I don't know. No volume of audit logging on the retrieval side changes what happens on the generation side, because governing access and governing truth are separate jobs that happen to share a wire. They always were.

What kind of knowledge never makes it into any system at all?

Start with the obvious category and then keep walking, because the obvious one is not the costly one. Everybody already knows the architecture diagram is out of date, and everybody has made peace with it. Fine. The expensive category is the knowledge that never had a document to fall out of date in the first place, and once you start looking for it, it has recognisable shapes.

The exception explained out loud, once, to whoever happened to be standing there. The workaround that worked, so nobody filed a ticket, because tickets are for things that are broken. The reason a rule exists, which almost never travels alongside the rule itself: a team inherits "run the batch twice at quarter end" and loses the partner defect that made it necessary in the first place. The veto that never happened, where somebody said no in a corridor and a bad idea died quietly, leaving no artifact anywhere a retrieval system could ever find it. The promise made on a customer call in March that three engineers are now building against without knowing it was ever a promise. Nothing filed.

The scale of that is measured now, and the measurement is bleak. APQC, with eGain as sponsor, surveyed 1,000 respondents in September 2025 and found that only eight percent of organizations consistently capture knowledge from departing retirees, which leaves 92% that do not, while 58% of C-suite leaders rated knowledge loss as strongly concerning or mission critical. Concern is not capture. Deloitte Insights picked the finding up on June 18, 2026 in an essay by Eyal Cahana and Evan Siegel that reframes the retirement wave as an institutional memory problem rather than a headcount one, and puts the point plainly: institutional knowledge lives in people, not systems, in the engineer who understands why a system was designed the way it was.

Put the two numbers beside each other and the picture resolves. Ninety-two percent of organizations are not capturing what their departing experts know, and the strongest available agent manages a bit better than a coin flip at recognising when it should decline to answer. Not two problems. That is one problem with a very fast delivery mechanism bolted onto it, and the delivery mechanism is the only half of the pair that improved this year.

How does structured discovery surface tacit knowledge before an agent needs it?

By treating discovery as an act rather than a query. A retrieval call asks a repository what it contains, which is a fundamentally different verb from asking a person what they know, and the second one has to happen before the agent needs the answer rather than at the moment the answer turns out to be missing. Timing matters most.

The mechanics matter more than the label does. One analyst interviewing one stakeholder produces one framing of the problem, filtered through whatever that analyst already believed walking in, which is how a workaround gets recorded as a requirement and a hard constraint gets recorded as a preference. One lens. Several specialists interrogating the same material from deliberately different positions produce something sturdier, because the questions collide and the collisions are where the tacit material surfaces. Specira runs five agents: four analysts reading the interviews, tickets and policies from distinct expert angles, plus a Red Team Critic whose only job is to attack what the other four produced. The critic is the part that matters for this argument. It is the one that asks why a rule exists rather than what it says.

Three questions do most of the work, and a repository can answer none of them. Who is the last person here who understands why this behaves this way? What did we decide not to build, and why? Which of these rules would surprise a competent stranger reading them cold? Ask those while the expert is still on the payroll and still has the patience for a long conversation, not in the fortnight after the farewell cake. Not after. That is the whole discipline, and it is less glamorous than the tooling.

Then connect the agents. Genuinely connect them, use the governed pipes the vendors spent this year building, because doing the discovery work and then stranding the output in a document nobody queries would be its own species of waste. The sequence is the argument, not the tooling. Capture what only people know, write it into the systems the context layer already serves, and the same MCP server that would have confidently guessed at that corridor workaround will instead retrieve it, cite it, and log the retrieval like everything else.

Nicolas Payette

Founder and CEO of Specira AI. 25 years in enterprise software delivery, most of it spent watching good systems answer confidently from incomplete context. Specira works the layer above the code, where the reason behind a rule is still a conversation somebody has to have.