Can the AI agent actually finish the work now?

The slower thinking

The thing that quietly changed

Most operators didn't notice that something changed this week. The news cycle was about Google's new search box and Gemini's new model, which is the consumer-facing layer of a deeper shift. The buyer-relevant news landed at the same time but in a different room. An engineer at Texas Instruments, Antoine Zambelli, published an open-source toolkit — toolkit meaning the engineering layer that sits between an AI model and the work it produces — along with a peer-reviewed paper accepted at a computer-science conference at the end of the month. The number from his eval is the one that matters: a small, free, generic AI model went from finishing 53% of multi-step tasks to finishing 99.3%, with no change to the model itself. The system around the model did the work.

That's the engineering version. The operator version is shorter. The reason your agent dropped the third step last year was not that the model was dumb. It was that the wrapper around the model — the part that catches a missed call, retries a failed lookup, refuses to send the brief without the regulator section — was missing. That wrapper has now been written, tested across ninety-seven model-and-backend configurations, and made publicly available under an open-source license. Vendors who were saying "the agent is reliable now" in 2023 were guessing. Vendors saying it in 2026 are pointing at a thing you can read and reproduce.

What changed is not the model. It is what gets wrapped around the model. And what gets wrapped around the model is what makes the brief land at 7am Monday with the regulator section included, the same way, every Monday.

Capability is not the same problem as reliability

The slower thinking starts by separating two things the marketing decks have been mashing together for three years: capability and reliability. They are not the same problem, they are not on the same schedule, and they are not the same vendor's job to solve.

Capability is what the model can do at all. Can it draft a brief in the voice of your CMO? Can it read a regulator's PDF and surface the three lines that bind your work next quarter? Can it look at the team's calendar and decide which Monday meetings need a one-page brief by 7am? Capability is decided by training — by the people inside the AI labs whose names you read about in industry-news posts about who joined where. Capability has been good enough for most operator work for about a year.

Reliability is whether the model finishes the work the same way, on the same kind of input, every Monday at 7am, for a year. Reliability is not decided by training. It is decided by the wrapper around the model — the part that runs an eval set every Sunday night to check that last week's outputs still look right, the part that refuses to send the brief if any of the five inputs failed to load, the part that retries a failed step with a different prompt before giving up. The wrapper is what every honest agent vendor in 2026 is selling, even if the marketing copy makes it sound like the model. The wrapper is the difference between "the agent answered when you asked it" (which has worked for three years) and "the brief is in your inbox by Monday morning with the regulator section included" (which is the part that has actually moved).

The reason this distinction matters to a buyer is that capability gets better on the AI lab's schedule, and reliability gets better on the agent vendor's schedule. If you're trying to figure out which work an agent can take, the question is not "is the model smart enough." The model has been smart enough for a year. The question is whether the vendor has built the wrapper, and whether they can show you the wrapper running on a recurring task. That's a much more concrete thing to ask for than "is your AI good." It's also a more discriminating one — most of the vendors with the loudest copy don't have a wrapper, or have one they can't show you outside a demo.

A worked example: the Monday-morning brief

Concretize this. Take a recurring task that almost every senior operator has on the calendar: the Monday-morning briefing that the team relies on to start the week. A small set of versions, all real:

• The COO of a fifty-person SaaS has been doing the Monday-morning team brief manually for two years. What changed in the market over the weekend, what the three competitors the team tracks shipped, what the regulator said on Friday afternoon, what last week's KPIs did, and what one decision the team is being asked to make by Wednesday. Two hours of Sunday-night work, every Sunday night, for two years.
• The CMO of a Series B consumer brand runs the weekly competitor monitor — five brands, three platforms each, plus the weekly TikTok / Instagram / YouTube creative monitor — and a Monday-morning report that goes to the agency, the founders, and the board's chair-of-the-marketing-committee. Three hours of Sunday-night work, every week, for eighteen months.
• The compliance lead of a regulated fintech runs the weekly regulator round-up — five named regulators across two jurisdictions, plus the trade-press monitor for any peer enforcement actions that might foreshadow theirs. Two hours of Sunday-night work, every week, plus a real-time monitor through the week.

The 2023 version of all three: the agent would summarize the first two inputs, drop the third, and produce something that was either wrong, but the operator caught it in time, or right, but missing the regulator section, or correct, but in the wrong voice. The operator went back to doing it manually. This is not a hypothetical. It is the experience that formed the inertia default this piece named at the top.

The 2026 version is the same five-step pipeline with the wrapper added. The agent runs on Sunday night, the eval set checks last week's outputs still pass before this week's run is allowed to publish, the brief arrives at 7am Monday with all five sections, and on the week the regulator's site is down the agent does not send a brief at all — it sends a Slack message saying one input failed to load, here's what I have, here's what I'm missing. The COO stops doing the brief. The CMO stops doing the monitor. The compliance lead keeps doing the judgment-call part of the regulator response — the part where two rules disagree and a person decides which one binds — and stops doing the scanning part.

This is, plainly, what FidelicAI does. The agents — KORA-01, VYRA-01, VEXA-01, and others by codename on the Roster — each take one of these recurring multi-step tasks. The wrapper is built in: eval set, step enforcement, retry logic, Slack-native failure mode, and a written document that defines the agent's scope — the deep-page detail of what each agent will refuse to do, surfaced when you read the agent's Roster page. The operator keeps the work that compounds — the strategy call, the customer in the room, the regulator-response-that-needs-a-name-on-it — and the recurring part shows up in Slack on Monday morning. The pricing is a small fraction of what a mid-market hire for that role costs. We make the case for the framing because the framing is what changed.

Where the news is right, and where it is narrative

Two honest corrections, because the rule of this register is to name what's narrative and what isn't.

The first: 99% on a benchmark is not 99% on your work. Zambelli's eval is run on a defined task surface — the kind of multi-step workflow the eval set covers. The day your input shape changes — the regulator files in a new format, your CMO changes the brief template, a competitor launches a feature whose name confuses the parser — the reliability number drops. The wrapper is not a button; it is a thing your vendor watches, re-evals, and re-tunes when the world shifts. The reliability problem moved from a cliff (the agent drops the third step) to a drift (the agent's eval score slowly decays as the world changes). That's a different problem, a better problem, and a manageable problem. It is not the same problem as "solved."

The second: open-source closes the technique gap, not the deployment gap. The fact that the wrapper exists publicly does not mean every agent vendor has built one. It means the technique is no longer secret, peer-reviewed evidence sits in the public record, and any vendor still saying "we have the magic sauce" is selling smoke. The right vendor question in 2026 is not "do you have a wrapper" — it's "show me your eval set, show me last quarter's drift report, show me what happens the week the input shape changes." The buyer who asks those three questions can tell within an hour which vendor has built the thing and which one is using the word.

The part to hold both at once: the engineering problem moved this week, in a way that is reproducible and public. The buyer's question — can I actually trust the agent to finish the work — has a different answer than it did in 2023. The follow-up question — who watches the wrapper when the world changes — is the new gating question, and it is a question about vendors, not about models.