# The coordination problem no agentic roadmap has priced

Niels Langereis, 20 July 2026. First published on LinkedIn: https://www.linkedin.com/pulse/coordination-problem-agentic-roadmap-has-priced-niels-langereis-ewrse

> Part three. Every company can rationally decide not to train the juniors it no longer needs, and the industry still ends up short of seniors in a decade. Nobody's roadmap has a line for the cost of that.

*Part three. The Financial Times has caught up to where this argument began. This is where it goes next.*

The [Financial Times](https://www.linkedin.com/company/financial-times/) ran a reassuring headline last week: AI isn't destroying entry-level jobs, it's changing them. The reassurance is not empty. US employers told NACE in April they expect to hire 5.6 per cent more graduates this year ( [National Association of Colleges and Employers](https://www.linkedin.com/company/naceorg/) , 2026 Job Outlook), and [PwC](https://www.linkedin.com/company/pwc/) , reading more than a billion job adverts, finds entry roles being steadily “seniorised” rather than deleted [(PwC, 2026 Global AI Jobs Barometer).](https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html) Two-thirds of the way in, the piece comes to the part that matters most: "strip out the mundane, repetitive tasks and you can starve early-career staff of the experience they will need when they eventually oversee hybrid teams of people and agents".

I made that argument in June, so I am of course gratified to see it confirmed by the FT. The mainstream has arrived at junior formation. Good. But arriving there is only the starting line, not the finish, and I want to say plainly what the next problem is, because it is more important than the one we just named.

A quick recap for anyone who missed the first two parts, and then I will move quickly past them. I argued that an AI-augmented system can inherit judgement, storing every senior call as precedent, without regenerating it in a single new person; and that you can rebuild the regeneration deliberately, with a second loop that routes real decisions back through juniors instead of letting the machine auto-resolve them. So, say we do that. Say it works, and you produce judges again. That still does not hand you a functioning organisation. This is the point [Patrik Eriksson](https://www.linkedin.com/in/patrik-eriksson-1118831/) put to me in the thread on the last piece, sharp enough that it deserves its own article.

More judgement is not more coherence. An organisation does not run on individual good calls; it runs on those calls adding up to one direction. The scarce skill, the one that does not automate, is not making a defensible decision. It is reconciling many defensible decisions into a single coherent one the organisation can act on. Call it coordination. It was always part of management. Agents are about to make it most of the job.

## Why agents make coordination harder, not easier

The intuitive feeling is that agents reduce coordination, because the machine handles the handoffs. The opposite is true. Every capable agent is one more confident, fluent producer of defensible calls, each optimised for its own local objective. Multiply the producers and you multiply the number of divergent, well-argued answers that have to be reconciled before anyone can move.

> i.e. the work of coordination does not scale with the number of decisions; it scales with the number of ways those decisions can conflict, which rises far faster.

And the raw material is now harder to adjudicate. I wrote last time about the epistemic gap that [Quattrociocchi, Capraro and Perc](https://arxiv.org/abs/2512.19466) mapped: models terminate in fluent, forced confidence, skipping the metacognition and value-weighing that a human decision passes through. Coordination has always meant weighing one expert's judgement against another's. Now several of the experts are systems that present every answer, right or wrong, in the same assured register. You are no longer choosing between a strong view and a visible hedge. You are choosing between two equally confident outputs with no tell.

## What coordination actually is

It is worth being precise, because the word 'coordination' feels 'soft'. **Coordination is not consensus, not averaging, not another meeting. It is the authority, backed by earned judgement, to adjudicate between competing defensible views, to hold the trade-offs that genuinely cannot all be met, to sequence them, and to own the resulting call when it turns out wrong.** It is a different faculty from making a single good decision, in the way that conducting is a different skill from playing an instrument well. We have spent two years automating the instruments. We have said almost nothing about who conducts.

## The coherence nobody owns

The structural shape is familiar if you read the shortlist piece. Every agent, every team, every function optimises its own objective, rationally and in good faith. Almost no one is measured on whether those objectives still add up to a coherent whole. Coherence is a public good inside the organisation, and like most public goods it has no natural owner, no line in anyone's targets, and no budget until it visibly fails. Local optimisation quietly erodes it, in the same way that agent-optimised discovery erodes the range of brands a person is ever shown. The failure is not loud. It is a slow drift into a company that is individually optimised everywhere and collectively incoherent, noticed only when the decisions stop pointing the same way.

This is where [Thomas Marzano](https://www.linkedin.com/in/thomasmarzano/)'s work earns its place in an operating conversation, not only a branding one. His argument that human value migrates upstream to constitutional judgement is, read properly, an argument about coordination: a constitution is the shared objective function that lets many autonomous judgers cohere without a person refereeing every call.

He is right that the constitution is the artefact that matters. But a constitution is a document, and coordination is a person applying it under pressure to a case it never anticipated. Writing the constitution is not the same as staffing the bench that runs it, and it is the bench I am worried about.

## Where does the coordinator get formed

Which lands us back where the series started, one level up. The 'second-loop' tells us how to form judges once execution is automated. It says nothing about how to form coordinators, and that is a scarcer thing to make. You learn to coordinate by owning a messy, cross-functional call while competing experts pull against you, being accountable for the outcome, and then doing it again. That seat, the one where a person carries a decision they did not fully control, is precisely the seat we are thinning out in the name of efficiency. We are on a course to regenerate judgement and still have no one who has learned to hold it together.

> So the question on the agentic roadmap is not only who are we still forming. It is who is accountable for the coherence of everything our judges, human and synthetic, decide, and where does that person learn the job.

**Right now that seat has no name on the org chart and no metric beside it. Most organisations will discover it was empty to begin with.**

## Sources

Andrew Hill, “AI isn't destroying entry-level jobs. It's changing them,” Financial Times, July 2026. [**https://giftarticle.ft.com/giftarticle/actions/redeem/45734087-ea6f-428c-bde8-aaa96ea76224**](https://giftarticle.ft.com/giftarticle/actions/redeem/45734087-ea6f-428c-bde8-aaa96ea76224)NACE, 2026 Job Outlook (5,6% meer graduate hires): [https://www.naceweb.org/talent-acquisition/trends-and-predictions/employers-expect-to-hire-5-point-6-percent-more-new-college-graduates-this-year](https://www.naceweb.org/talent-acquisition/trends-and-predictions/employers-expect-to-hire-5-point-6-percent-more-new-college-graduates-this-year)PwC, 2026 Global AI Jobs Barometer (seniorisation, 1bn+ adverts): [https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html](https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html)Quattrociocchi, Capraro & Perc, "Epistemological Fault Lines Between Human and Artificial Intelligence," arXiv:2512.19466 (22 dec 2025, titel/auteurs/Epistemia-concept bevestigd): [https://arxiv.org/abs/2512.19466](https://arxiv.org/abs/2512.19466)Thomas Marzano, "Brand Constitutions: The Legible-Lovable Standard," Brandingmag: [https://www.brandingmag.com/product/brand-constitutions--the-legible-lovable-standard-for-building-equity-in-an-agentic-economy/213/](https://www.brandingmag.com/product/brand-constitutions--the-legible-lovable-standard-for-building-equity-in-an-agentic-economy/213/)Deel 1, "If AI takes the junior work, where does the next senior come from?": [https://www.linkedin.com/pulse/ai-takes-junior-work-where-does-next-senior-come-from-niels-langereis-ffw1e/](https://www.linkedin.com/pulse/ai-takes-junior-work-where-does-next-senior-come-from-niels-langereis-ffw1e/)Deel 2, "Juniors don't build judgement by overseeing AI": [https://www.linkedin.com/pulse/juniors-dont-build-judgement-overseeing-ai-niels-langereis-hqxoe/](https://www.linkedin.com/pulse/juniors-dont-build-judgement-overseeing-ai-niels-langereis-hqxoe/)