The Agentic Org Chart: Team Shape When Your Teammates Are Agents

Most of the conversation about AI and engineering teams is really about roles — what the humans now do once agents take the typing. That question matters, and we treat it elsewhere. This is the other half: not who does what, but how the team is shaped. When some of your teammates are autonomous agents, what is the right size of a team, where do its boundaries fall, and does the old wisdom about org design still hold? The intuitive answer — agents dissolve team structure, one engineer plus a fleet replaces the team — turns out to be wrong, and the evidence is clearer than the hype.

Team Topologies survive — because software is still intent as code

The authors of Team Topologies, Matthew Skelton and Manuel Pais, have been explicit that the framework’s fundamentals do not change under AI, because the thing being produced has not changed: software is still knowledge work that represents intent as code. Cognitive load, clear domain boundaries, and the flow of value through a team are still the constraints that govern good design. Their sharper claim is that AI does not remove socio-technical complexity — it amplifies it, which makes deliberate team design more important, not less. This lines up with the DORA finding we keep returning to: AI is an amplifier of the system it lands in. Drop agents into a team with muddy boundaries and you get faster mud.

The most counter-intuitive survivor is the size limit. The roughly eight-person ceiling on a team, rooted in the human limits Dunbar described, persists — agents do not abolish it. What changes is what a team of that size can take on: a small, bounded team can now own a larger domain than before, because agents absorb more of the execution. We checked the stronger versions of this claim and they did not survive scrutiny — that Dunbar stops applying, that teams become majority-agent, that a lone engineer owns an entire product. The durable statement is narrower and more useful: the human side of the team stays small and bounded; the domain it can hold grows.

Span of control: the unit becomes a human plus a fleet

The real structural change is at the level of the individual. The unit of delivery shifts from a person doing the work to a person orchestrating a fleet of agents doing it — broadening one engineer’s span of control from a single stream to several at once. Oversight moves from human-in-the-loop, where a person touches every step, to human-on-the-loop, where a person supervises the system and intervenes where it matters. But the span has a ceiling of its own, and it is set by the scarce resource we name across this whole body of work: acceptance. One engineer can orchestrate as many agents as they can still meaningfully review and answer for — no more. Push past that and you have not increased capacity; you have created unowned change moving at machine speed.

Agents widen one engineer’s span of control until it hits the only hard limit that matters: how much change a human can still review and be accountable for.

Conway’s Law when teammates don’t obey social limits

Conway’s Law — systems come to mirror the communication structures of the organisations that build them — still applies, but its premise shifts in an interesting way. Part of the communication structure is now made of agents, and agents do not have the social bandwidth limits that constrain human teams; they coordinate through interfaces, contracts and context at a scale no human org could. That makes the inverse-Conway move — shaping teams and boundaries to promote the architecture you want — more powerful, not less. The organisations best positioned to adopt agents are the ones that already practise what amounts to bounded agency for humans: clear ownership, stable interfaces, least-privilege access, and traceability. Those same properties are exactly what an agent needs to act safely inside a domain. Good human boundaries turn out to be good agent boundaries.

There is a real cautionary side to this, and pretending otherwise would be dishonest. A controlled study by METR found experienced developers were slower, not faster, with early-2025 AI while believing the opposite — the new bottleneck is human review throughput, not generation. Gartner expects more than 40% of agentic-AI projects to be cancelled by end-2027. The lesson is not to avoid agents; it is that the org design — the boundaries, the oversight, the accountability — is a first-order part of making them work, not an afterthought to the tooling.

So the agentic org chart is less of a redrawing than people expect. Keep teams small and their domains bounded; design the boundaries deliberately, because agents will encode them whether or not you meant them to; and treat the human-plus-fleet, with its acceptance-limited span of control, as the new unit of delivery. For the roles those humans now hold inside this structure, read the AI-native engineering team; for why the boundaries matter at all, why alignment beats agile.