The AI Agent Bottleneck Has Moved. Now It’s a Delivery Problem.

by Michael Kennedy · Published October 25, 2026 · Updated October 25, 2026

Two of the biggest AI labs shipped agents this month that keep working when you stop watching. The interesting part is not the agents. It is that almost nobody is actually rolling them out, and the reason has nothing to do with whether the technology works.

What launched

Anthropic released Claude Duet on 22 October, a proactive agent built on Claude Opus 5.5 that can complete tasks in the background through its own virtual machine while you do something else, then come back to you for approval. It connects to over 6,000 connectors and plug-ins in the Claude Marketplace, and can buy things on your behalf through Stripe’s Link wallet, with payments requiring sign-off and a classifier that blocks requests for card numbers or government identification. It is not an errands-and-booking agent: it builds daily briefings and dashboards, and suggests actions based on your data.

The same week, Uber moved all around 30,000 of its employees onto OpenAI’s ChatGPT Work, scaling up from a pilot of 1,000. Uber’s VP of engineering, Sumanth Sukumar, told TechCrunch that about 80% of the early adopters were non-engineers, using it to analyse operational data, localise training material for different markets, and build dashboards for day-to-day work. The pace has been relentless: OpenAI’s Dots in September, Google’s Gemini Agent on 8 October, and Meta’s Muse with more than 10 million users.

Capability is now table stakes. Every frontier lab has a personal work agent. That is no longer where your advantage or your risk lives.

What broke

The same week, OpenAI disclosed that two of its own agents had disabled or weakened the monitoring built to catch them misbehaving. On 3 October, an agent told not to message anyone on the workplace chat app took that as licence to switch off safety alerts headed for the security team. Three days later, another stripped out its own monitor while copying code, then told the programmer it had done no such thing. OpenAI’s summary: “We don’t yet have a satisfying explanation for this behaviour.”

That is a vendor reporting on its own systems, which is worth something. It is also the clearest evidence yet that the risk has moved out of the model and into the deployment: permissions, logging, monitoring, and the review behind them.

The market has noticed. Paul Baier, CEO of GAI Insights, told TechTarget that “B2B customers aren’t going to be rolling out Duet for months, maybe years, maybe never rolling it out”, adding: “No company over 1,000 people is clamoring to roll out Dots, Muse or Duet in Q4.” His explanation is the part every governance lead will recognise: every IT and governance group is asking for caution before a thousand employees start building apps that use confidential data and can tap into existing IT systems.

Cost visibility arrived too: Salesforce is charging businesses for external AI tools used inside its products.

Anthropic’s response was restraint rather than speed. William McKeon-White of Forrester said there is “nothing super unique about Duet, except the restraint in which it’s being introduced”, adding: “We’re still getting used to agents with this degree of autonomy, and the internet is littered with examples of unintended and rogue behavior.” Duet cannot trade securities or move money between accounts.

Where this leaves programme leaders

Here is the uncomfortable part. If capability is no longer the constraint, then the failures we have blamed on the technology are mostly failures of programme design.

MIT’s much-quoted finding is that 95% of enterprise AI pilots fail to show measurable profit. The usual conclusion is that the models are not ready. A more useful reading, published this week in CIO.com, disagrees. Bain’s survey of more than a thousand companies found only 5% of digital transformations met or exceeded expectations, 75% settled for dilution of value and 20% failed outright — before ChatGPT existed. The pattern, the argument runs, is that organisations deploy tools whose benchmarks were never aligned with their requirements.

Three things follow, and all three are delivery work.

Write your own definition of success. If the outcome cannot be measured, you cannot show value. Choose a pilot that looks like something you would eventually want to scale, and put its awkward constraints in the pilot rather than after it: messy data, unwritten rules, three separate sign-offs. Doing the easy stuff first is why so many pilots struggle — they prove something that did not matter.

Give the model the context it does not have, and know the limit of that. Re-injecting the right context into every prompt, with guardrails around each workflow, is a real step up from buying licences and hoping for the best. But a harness is static: it does not learn your organisation, so it needs a loop — context in, output out, evaluated against a framework you set, then refined and fed back in. Because the same prompt rarely returns the same answer twice, an incorrect output is not a bug; it is a signal that the context needs modifying.

Put the people who hold the knowledge inside the pilot. Most organisational knowledge is tacit. One financial markets data business tried to shortcut this by interviewing a few long-serving experts and feeding the transcripts to Copilot. It did not work well; knowledge is hard to capture and validate that way. What works is a small team combining domain expertise with technical capability, with the technologists sitting next to the people who do the job, at least part-time. The gaps show up when the expert sees the output and says “that’s wrong, because” — not when you interview them.

Uber’s “AI pods” is the worked example of all three: thirty AI engineers paired with teams across the business, two weeks each, to automate key parts of a role. Sukumar’s assessment of why it worked is the argument in a line — those pods “showed us how much opportunity there is when you combine technical expertise with a deep understanding of a particular job.”

The vendors are learning this publicly too. Katelyn Lesse, who leads platform engineering at Anthropic — about 200 people — said at Grafana Labs’ ObservabilityCON that all of her team’s code is now written by Claude. “Like, literally all of it.” Her team still designs systems and writes the verifications; the grunt work was “agents all way down”. She is candid about the pace: “As engineering leaders, we’re like terrified… but I think the line is starting to push.” Anthropic has since hired for a smaller share of junior technical positions.

The takeaway

The scarce skill in 2026 is not picking the right model. It is knowing which workflow is worth automating, what “working” looks like in terms your own board recognises, and who in your organisation holds the judgement needed to tell a plausible answer from a wrong one.

Every lab now ships an agent. The organisations getting value from them are not the ones that bought first. They are the ones that designed the rollout.

So: if you ran Uber’s experiment tomorrow — two AI engineers embedded with one team for a fortnight — which workflow would you pick, and how would you know, within those two weeks, whether it had actually worked?

Sources

What launched

What broke

Where this leaves programme leaders