The Extra Chair
Why adding more people can leave your company slower, even when everyone does their job well.

You were faster when you were smaller. Now there are more capable people, and getting anything through takes longer. You add people to share the work, then process to keep them aligned.
The newest answer is agents: help without hiring or meetings. You add them alongside the people and process, and more work counts as finished, but you still wait to ship.
That can make the failure feel personal. You know the people haven't stopped caring. The effort that once produced progress now goes into keeping everyone informed.
Growth raises four costs together: coordination, lost context, late feedback and the structure's limit; more people need to agree, summaries leave things out, and results arrive after another decision. The approval route can keep all that effort waiting.
The Momentum Engine is these four costs working together, and adding approval to catch a mistake, for example, can delay feedback. Agents added to an unchanged arrangement bring these costs too.
Consider a thought experiment: a team adds five agents to one campaign, each writing its own part: the copy, the bids, the landing page, the email and the report.
The extra chair
The extra chair means another possible conversation with everyone already involved. Five people have ten possible pairwise connections; ten have 45. Fifty have 1,225 possible pathways, though that isn't a count of conversations everyone must have.
Some work divides easily. Other work requires each person's decisions to fit everyone else's. The second kind makes more connections necessary before anyone can finish.
The campaign has six chairs: five agents and the person who owns it; that makes 15 possible pathways, and the copy, landing page and email must fit together.
Fred Brooks described that condition in The Mythical Man-Month in 1975. He had managed IBM's OS/360 project, involving more than 1,000 people at its peak. His warning was: "Adding manpower to a late software project makes it later."
Brooks called that an oversimplification. His distinction was between divisible work and work whose parts must fit. Adding people can help the former while burdening the latter.
Anthropic described its limit in 2025. Tasks requiring shared context or many dependencies weren't good fits for multi-agent systems then. Its researchers also reported that multi-agent systems used about 15 times the tokens of chats.
Brooks's condition concerned people; Anthropic's concerned agents. The dependency matters, not the meetings; the parts still have to fit.
Kelly Johnson treated that cost as something to contain. His Lockheed team designed and built the XP-80 in 143 days, and his rules required project headcount to be "restricted in an almost vicious manner".
The telephone
Fewer participants won't help if the message loses its meaning. The telephone is meaning lost as work passes between people, like the children's game, and a summary can preserve the decision while dropping why it made sense.
Suetonius describes Augustus first placing young men at intervals along military roads. They passed reports along the route. He later used carriages so the same courier could bring a dispatch from its origin.
Augustus could then question the person who had brought it. Historian Anne Kolb explains that the relay's final messenger lacked personal knowledge of the contents. He could deliver the message but couldn't explain its background.
In the campaign, the angle the copy agent chooses never reaches the landing-page agent. Walden Yan of Cognition described the principle in 2025.
Each has made choices the other can't see. Yan's principle was: "Actions carry implicit decisions, and conflicting decisions carry bad results". The task description doesn't carry every decision needed to make the parts agree.
A handoff inside a company can lose that knowledge too. Augustus kept someone available who could explain the message.
The lag
A courier who can answer questions can still arrive late. In 1959, Milton Friedman warned that policies "operate with a long lag and with a lag that varies widely". The uncertainty concerned timing as well as delay.
In A Program for Monetary Stability, he reported money-growth leads of six to 29 months at business peaks. At troughs, they ranged from four to 22 months. A company isn't an economy, but it can also act before learning what the last action did.
In DORA's 2018 survey, elite software teams reported deploying on demand, multiple times daily. Their time from committed code to production was under an hour. DORA calculated 46 times more frequent deployment, 2,555 times faster lead times and seven times lower failure rates than low performers.
Those multiples came from normalized survey ranges, not measured averages. The speed paradox means going faster came with fewer failures, not more.
In the campaign, the five agents produce faster than the owner can learn what the last change did. In 2026, Yan reported that teams using many agents became constrained by management, planning and reviewing. Faster generation moves the delay into review when nobody changes that arrangement.
Production gets ahead of learning. Short loops help people find errors before more work depends on them.
The ceiling
Gothic builders couldn't keep raising walls without changing their support. Thicker walls left less room for windows. Flying buttresses carried the outward thrust outside the wall, letting it be thinner; Notre-Dame's nave rises 33 meters.

Moving the load changed what builders could do with stone. In 1968, Melvin Conway argued that organizations produce designs constrained by how their members communicate. The resulting systems become "copies of the communication structures of these organizations".
Early well-known multi-agent systems explicitly copied company arrangements. ChatDev described a virtual software company with a CEO, CTO and programmer.
MetaGPT assigned product managers, architects and engineers to an assembly line. The campaign follows the same pattern: its agents are split like departments, into copy, media, web, email and reporting.
Neither description proves those systems failed. They show that an agent arrangement can inherit familiar roles and handoffs, and the dependencies matter even when the roles run cheaply.
A structural zombie is an old route that outlived its use but still directs work. A new org chart can leave decisions following the same approvals. The labels change while the waiting remains.
When the fix adds to the cost
Hire to go faster, and more people need to coordinate. Specialize their roles, and more work crosses boundaries; add approval to catch what gets lost, and feedback takes longer.
Each addition can have a sensible local reason. The overloaded person needs help, and the specialist brings useful skill. Neither reason tells you what the addition costs elsewhere.
This is the treadmill: hiring to go faster and needing more coordination, specializing roles and crossing more boundaries, adding approval and slowing feedback. Each response to the strain adds to the cost producing it. The fix becomes another source of friction.
In the campaign, add a sixth agent to check the five, and the review queue grows. The extra check may catch mistakes while delaying what reaches the customer.
Relief still matters during a live incident. Sometimes capacity really is the constraint, without an upstream defect producing the shortage. Stopping damage or adding useful help can be the right decision.
Relief needs a follow-up when work keeps arriving broken. Fixing one link can load another: faster production can overwhelm review. The useful test is whether the whole route improves, not whether its busiest point produces more.
The customer gets the whole route
The customer pays for those internal costs. A site can mirror departments while each department considers its part complete. The customer still has to find a way through it.
Jakob Nielsen warned in 1997: "Users should not have to care how your company is organized". Nigel Bevan likewise described websites reflecting internal concerns rather than users' needs. A structure sensible to the company can make the customer work harder.
Empathy Inversion means building a message for the room where it was written, rather than its user. A website organized around departments is one example. The customer meets the communication structure as the product, and the delay as waiting.
Early in my career, I ran the same SEO for two clients and got the same rankings. One grew. The other saw no return it could trace.
The channel report was right both times. It couldn't account for how the rest of each business affected the result.
An agent's correct draft is like that channel report: right about its own part and silent about what the customer got. A report that is right but can't explain the result isn't ready to reach the client.
A department boundary can mark where one team's part ended. It doesn't change what the customer got. So I judge the work by what the customer got, and instead of a report, I ship the change that closes the gap.
The chair that costs almost nothing
I expect AI agents to make the cost harder to see. More precisely, I expect their low starting cost to make it easier to ignore. Anthropic's token figures already show that the added chair isn't free.
Research first posted in December 2025 found sharply different results by task: Yubin Kim and colleagues reported multi-agent losses of 39 to 70 percent on sequential planning. Gains reached about 81 percent on parallel financial reasoning, compared with a single agent.
Kim and Xin Liu wrote that "the overhead of communication fragmented the reasoning process"; independent agents amplified errors 17.2 times, against 4.4 times with a central coordinator. In the campaign, five agents each writing their own part are Kim's independent agents (17.2 times); one owner holding the writes plays the central coordinator's role (4.4 times).
The copy, landing page and email must fit in order, the kind of work where Kim found the losses. I read that as reason to examine dependencies and context lost at handoffs before counting agents.
DORA's 2024 survey associated each 25 percent rise in AI adoption with lower delivery performance. Estimated throughput fell 1.5 percent and stability fell 7.2 percent. These were survey-model associations, not proof of cause.
The 2025 findings changed part of that picture, as adoption now went with higher throughput, while instability still rose. DORA described AI as amplifying existing strengths and weaknesses.
My expectation is that agents added to unchanged structures will raise output and breakage together. I'd be wrong if DORA's next data showed teams adopting agents without changing team shape, batch size or handoffs, while stability rose alongside throughput. That would give me reason to revise the expectation.
One writer per route
The coordinator result favors one owner: the five agents produce faster than the owner can review, so more review isn't the fix. The fix is fewer writers.
One writer per route means one owner controls changes reaching the customer. Everyone else can read, research and propose.
Ten months after his first post, Yan reported what was working; in April 2026, he described arrangements where "writes stay single-threaded". Other agents contributed intelligence rather than independently changing things.
The working shape had a manager split the work. Child agents executed their tasks, and the manager combined their findings and reported back.
In the second setup, four helpers read, research and draft; one owner holds every write reaching the customer. That owner judges the parts together before anything changes.
Enforce that distinction through permissions. Helper agents may change drafts, analysis and proposals, but not the score, the customer record or the send. Hold every write that reaches a customer until the route's owner sends it, changes it or stops it.
Check the campaign's customer records for what each buyer does next on the route the write touches: a reply, a booked call, a purchase or an unsubscribe. If those do not move, the change fails, however much the agents produce.
Marketing and sales examine those same records rather than treat separate reports as the result. The Market-Side Builder owns the route all the way to the customer and holds its writes, with authority to ship the change closing the gap.
Give each route its own owner, rather than send every route to one company-wide reviewer. Otherwise, the permission check recreates the central approval queue.
Large doesn't have to mean slow
Good people and better tools can reduce these costs. Async communication makes some connections cheaper. Yan also reported that models had become more capable of acting as agents.
Better models improve what each helper contributes. Nor must every possible connection become a real conversation, because structure determines which connections the work requires.
Netflix reported thousands of deployments daily by 2018, and by 2021, it reported more than 20,000. Size clearly doesn't make speed impossible.
Amazon's two-pizza teams own a specific product or service. Small size matters alongside ownership that lets work finish without company-wide agreement. A smaller team still waiting on central approval lacks that independence.
Copying team names isn't enough. Former Spotify product manager Jeremiah Lee said its squad model "was only ever aspirational and never fully implemented". Yan's remaining agent problems also concern coordination: "Agents assume they share state with their children when they don't."
Cheaper communication helps. It doesn't remove the need to decide who knows what and who may act.
Small where the work happens
Buurtzorg changed those decisions in Dutch home nursing. Jos de Blok, a district nurse who had also worked as a manager, founded it in 2006. Small teams took responsibility for care in their neighborhoods.
Teams have at most 12 members and split when they grow, while coaches help without authority over them. A back office of about 50 supports more than 14,000 staff.
The extra chair is contained by the team-size limit, and the telephone is shortened because one team owns the patient's care. The lag is shortened because the nurses doing the work see its result.
The ceiling changes because teams needn't wait for middle management to direct care, and the arrangement changes who holds the context and who can act.
In the campaign, one owner per route contains the chairs. The owner sees every part, so the telephone shortens, and judges by what buyers do, so the lag shortens. Without a central approval queue, the ceiling moves.
KPMG's 2015 study, using 2013 data, found 108 home-care hours per client yearly, against 168 elsewhere. Satisfaction was among the highest, but total cost was about average after follow-up care. Quality came from patient surveys, not clinical outcomes.
Listen to the frustration before choosing the remedy: "We were faster when we were smaller" points toward coordination. On an agent route, count the writers who must agree. "It got lost in translation" points toward handoffs: check whether the landing-page agent can see why the copy changed.
"We don't know what's happening" points toward feedback. Ask whether the owner sees buyer responses before the next write. "The system won't let us" points toward the ceiling: look for a central approval queue.
Before adding a person or agent, ask who owns the route and who holds the write. You were faster when you were smaller, but that doesn't mean the whole company must shrink. You may need less of it involved in each decision.
Growth doesn't require every task to grow with you.
I keep recording where the work slows, as I find it.