Memo #41: The Knowledge Handoff Problem already established the claim this memo prices, not the one it discovers: “The Coordination Tax does not disappear when humans are replaced by agents.” It migrates to the agent boundary and resurfaces there as Context Collision. A workflow with five agents in sequence has four internal handoff points. Each one is a potential Context Collision event. What #41 does not answer is the question that follows from its own math: given that every additional agent is another handoff, how many can a task afford, and where does adding one stop paying for itself?
Memo #116: Fast Is Not the Same as Profitable gave the unit. Cost per Completed Outcome is the fully loaded cost of one resolved unit of output — compute, Steward time, and residual Coordination Tax — divided by outcomes actually completed, reported by Task Tier. An added agent earns its place only if it lowers that unit at the tier it sits in, or protects Revenue per Steward by removing Steward load. Architecture diagrams with more boxes are not evidence of either.
The tax has already been shown to survive three transitions: human approval chains after task automation, structural redesign that leaves escalation paths intact, and fully agentic handoff. This memo does not prove the cost survives again. It asks which additions are worth it.
What a boundary costs even when it is built correctly
Memo #41’s remedy for Context Collision is shared operational state — a governed knowledge layer with defined write permissions, versioned entries, and a Proof of Action trail. That remedy works. It is also a standing cost, not a one-time fix, and the marginal cost of the fourth handoff is not the cost of standing the ledger up for the first time.
A correctly designed handoff still carries five costs every time it executes. The categories recur at every tier. The magnitudes do not.
Context transfer — the reasoning package has to be written, versioned, and read. A T1 payload is small. A T3 payload carries exception history and confidence. Latency — delay on the critical path, not idle calendar time. Inference — each agent is its own model call, unless the hop exists only to route the step onto a cheaper model. Failure surface — more places for Handoff Friction or Context Collision, each of which needs Deterministic Failure design or it fails silently. Accountability — Delegation Trace records which agent acted under which authority. Six writes are cheap. Steward time reading an unattributable trail after a failure is not — and that time lands in Cost per Completed Outcome and in MTTI.
Sequential hop-count is the conservative case, not the only architecture. Fan-out and a router-plus-specialists pattern can reduce total inference if they stop the wrong model running. Those shapes still add boundaries. They have to beat the same test. They are not automatically Theater, and they are not free.
The decision rule
Add the agent only if the cost it removes exceeds the cost it adds, measured in Cost per Completed Outcome at that tier.
What it adds: context write/read, critical-path latency, incremental inference, expected failure and rework at the new hop, and Steward time spent on trace review when the hop fails. What it removes: Steward minutes, reversals, and failed outcomes the previous design was paying for. If both sides cannot be stated, the hop is not priced. If it cannot be priced, it does not ship.
The Simulation case study on customer-care Workforce Arbitrage does not measure orchestration cost directly. It does give the magnitudes a boundary has to beat. Under that model, a completed T1 ticket in the Stewardship stack costs on the order of four euro cents; a T2 ticket about €1.30. Steward time at reference volume sits at 5% on T1, 35% on T2, 60% on T3. Those are modelled figures against published labour costs, not live operating results. They are enough to show why the default at T1 is not to add the agent, and why a review layer is the tier where a new boundary can still clear the bar.
Pricing the boundary by task tier
At T1, Intervention Threshold targets roughly 1:100 and Escalation Rate confirms it. There is little Coordination Tax left for an additional agent to remove. Splitting a T1 task across two agents adds a handoff’s full cost against a saving that is already close to zero. That is the default. It is not a prohibition. A cheap checker that prevents a rare expensive reversal can still beat the inequality. The test is the number, not a tier taboo.
At T2, Intervention Threshold runs 1:5 to 1:10, and the calculus flips. Agent Council is the architecture that shows why. The Council inserts specialised review agents between the Execution Layer and the human Judgment Layer, handling quality review and escalation triage at compute cost instead of human cost. That layer is itself an added boundary. It earns its cost by removing a more expensive boundary downstream: every case the Council resolves is a case that never reaches the Steward, and Steward time is the most expensive resource in the stack. This is the tier where adding a boundary is most likely to pay, because it trades a compute-cost handoff for a human-cost one.
At T3, mandatory human judgement does not move. An added agent cannot remove the terminal dependency on the Steward. It can only change what the Steward sees. A pre-processing boundary that hands over a cleaner exception can be worth its cost. A T3 layer that exists to look thorough before a decision it cannot replace is paying full handoff cost for no removable tax.
Orchestration Theater
Orchestration Theater is the pattern in which agents are added because more agents read as more sophisticated, without a pre-specified, measurable movement in Escalation Rate, MTTI, Cost per Completed Outcome, or Steward load. The workflow looks more advanced. Nothing about its economics has moved.
Delegation Trace is how Theater gets caught, not the definition of the crime. A five-agent chain with no way to attribute a specific outcome to a specific boundary cannot prove that any hop earned its cost. Missing instrumentation is how the pattern hides. The pattern itself is the added boundary that never named what it was supposed to remove.
The Operator’s Verdict
Multi-agent is not a synonym for more autonomous. The number of agents in a workflow is not a measure of anything. Every agent added is a boundary. Every boundary costs context transfer, latency, inference, failure surface, and Steward time on the accountability trail — the same five categories whether the boundary was worth adding or not, at magnitudes that follow the tier.
Price each addition against what it removes, in Cost per Completed Outcome, by tier, before it ships. A T2 boundary that trades Steward time for compute time has usually paid for itself before it runs its first case. A T1 or T3 boundary added because the diagram looked unfinished usually has not. The fact that it looks more sophisticated is not evidence to the contrary.
KEY TAKEAWAY
Why isn’t the Coordination Tax eliminated simply by replacing human coordinators with agents?
Because the tax is generated by the number of boundaries a task passes through, not by whether humans or agents sit at each boundary. Memo #41 established that removing humans relocates the tax to the agent-to-agent handoff, where it resurfaces as Context Collision. Memo #116 supplied the unit. Even a correctly designed handoff costs context transfer, latency, inference, failure surface, and Steward time on the Delegation Trace every time it runs. A boundary is only worth adding when it lowers Cost per Completed Outcome at that tier — as Agent Council does at T2, converting human-cost review into compute-cost review — not by default.
