The 4 Edges of Human-Agent Coordination: H2H, A2A, H2A, A2H in Hard Currency
Every hybrid decision crosses four types of edges. Each has a distinct cost, and each wins under different patterns. No one measures them separately, which is why the bill escapes.
90-Second Summary
Every big decision moving through your company today is a line that hops between boxes. Some boxes are people, others are AI agents. The cost does not live in the boxes, it lives in the lines: human coordinating human (H2H), agent coordinating agent (A2A), human coordinating agent (H2A), agent coordinating human (A2H). Four types, four different economies. Multi-agent architecture took off across 2025 and 2026, and most companies still cannot say, with any clarity, what each agent actually did in production. Add the four up into one number and you erase the most expensive distortion the AI wave shoved into your operation, ending up with a tidy total that tells you nothing about where to push.
Think back to the last decision of weight that crossed your desk with AI in the middle. A renewal for a big account, a price adjustment, a budget sign-off, a technical hire. Before it came out the other side as action, it ran through some three humans and two agents, and almost none of them logged the route.
It went roughly like this. An analyst asked Claude for a draft with three scenarios, got it back, and tuned the output across three rounds. Sent it to the head of finance, who opened Copilot, cross-checked it against the in-house numbers, and bounced two questions back to the analyst. In the end the two of them sat down with finance leadership to call it. One decision, three people, two agents, and four different kinds of edge crossed without anyone watching the clock.
Each of those edges charges in its own way, and each is the most expensive one in a different scenario. Because nobody measures them separately, the whole invoice slips under the radar.
The Modern Decision Graph Has Four Kinds of Edge
Call this network a graph if you want the technical name, or a map of boxes and lines, same thing. The nodes, the boxes, come in two types only. Type H, people: someone on the team, a contractor, an outside advisor. Type A, agent: a model with tool access, an agent that runs on its own, a system that decides automatically and takes the place of one more node. Put the two types together and the links between them give just four possible combinations. Four, no more. And it is across these four lines that the money drains away.
| Edge | Movement | Concrete example | Dominant cost | Who pays |
|---|---|---|---|---|
| H2H | Human coordinates human | Meeting to settle how the team uses Claude; senior signing off on the AI output; the specialist pulled in when the agent gets stuck. | Synchronous time × senior loaded hourly rate | Payroll of everyone in the room |
| A2A | Agent hands context to agent | Claude output reshaped to fit Copilot; the handoff between two agents; a format one pipeline outputs and the other will not swallow. | Tokens (low) + human fix when the handoff fails (high and quiet) | Senior engineering payroll on the fix |
| H2A | Human triggers agent | Analyst writes the prompt and refines it 4 to 6 times until the output is usable; ongoing tuning of the command. | Senior time × prompt rounds | Payroll of whoever operates |
| A2H | Agent delivers, human checks | Whoever signs off reads the AI output, checks the assumption, asks for a fix; rework when the agent gets a base assumption wrong. | Senior review time × chance of rework | Payroll of whoever signs off (the priciest) |
Each edge has its own economic behavior. Each one grows on a different trigger. Treating them as a single block is exactly what produces the hidden invoice that carries most of the cost of human-agent coordination in companies that adopted AI over the last 18 months.
H2H: Classic Human Coordination Did Not Vanish, It Got More Expensive
You let people go, you buy AI, you automate a process, and intuition promises the obvious relief: fewer people on payroll, fewer people to align. In the operation the opposite happens. Whoever is left after the AI-driven pruning tends to be more senior, more specialized, more expensive per hour. Payroll shrinks in headcount and the average hour of what remained goes up.
The H2H edge did not get rarer, it got more expensive per unit. That 90-minute meeting with four seniors to settle what comes out of the AI used to cost, on the old math, R$ 1,440. Today the same meeting runs R$ 1,920. Same clock, the bill went up a third, and the budget keeps carrying the old number as if nothing had changed.
And AI even invented a kind of meeting that did not exist before: the one whose entire agenda is "how the team is using AI". Group calibration, prompt alignment, the fight over which tool fits which case. Your company can run three a week and count none of them as a product or operations meeting, because they live in a corner of the calendar without an honest label. When the habit earns a fixed slot and minutes, it goes by the name of the AI committee, the typical anti-pattern of 2026.
| H2H sub-pattern | Typical frequency | Unit cost | Warning sign |
|---|---|---|---|
| Meeting to settle how the team uses AI | 1-3 per week | R$ 1,200 to R$ 2,800 | Shows up on the calendar with no clear owner |
| Senior signing off on the AI output | 4-8 per week, per area | R$ 480 to R$ 960 | Number of seniors did not fall, even with AI |
| Specialist pulled in when the agent gets stuck in production | 1-2 per week, per critical area | R$ 600 to R$ 1,600 | Senior engineering interrupted off the agenda |
| Committee to re-approve an automated decision | 1-2 per month, per risk area | R$ 2,500 to R$ 8,000 | Customer questioning the AI's logic |
| Prompt alignment between neighboring areas | 2-4 per week | R$ 800 to R$ 1,800 | AI output drifting between teams |
A2A: The Silence That Costs a Fortune
From a distance, A2A is the only edge that looks cheap. Token, API call, machine time: add it all up and you get pennies per interaction. Infra FinOps measures that part just fine, and that is where the trap sits. The real bill is not on the cloud invoice. It is on the senior pulled in to fix what two agents could not hand off to each other.
The format mismatch is the usual case. Claude returns nested JSON; the next agent only swallows a flat shape. Copilot sends 12 optional fields; the pipeline downstream counts on 8 mandatory ones. None of this blows up as a red error on the screen. It comes through as a half-delivery someone has to patch by hand, and that someone earns a senior salary.
Multi-agent architecture took off over the last two years in enterprise operations, and most companies still cannot say what each agent did in production. In a shop with four to six agents running across different pipelines, fixing these seams eats R$ 8,000 to R$ 15,000 a month of senior engineering payroll, and that number is written down nowhere.
There is a quieter signal, and it shows up on the technical team's calendar. The senior engineer is "lending a hand" to the non-technical areas at a frequency nobody added up. Marketing asked for help getting the AI output to fit the email template. Sales asked someone to clean up what Claude spat out before it went into the CRM. Each of those little favors is an A2A edge with the bill pushed onto a payroll that never expected to pay for it.
H2A: Prompt Iteration Is an Invisible Line on Payroll
H2A is where the human pokes the agent. The cycle you see in practice runs four to six prompt rounds until the output is usable, somewhere between 35 and 50 minutes. And whoever is in that back-and-forth is no intern: AI already ate the junior's work, so the person left refining the command is expensive. Per cycle, model it at R$ 220 to R$ 400, and redo it with your own loaded hour.
That bill grows in proportion to adoption. Say 40% of the team operating with AI in the flow, each person running about three H2A cycles a day: in a 500-person company, the H2A edge alone weighs R$ 250,000 to R$ 460,000 a month. It is money with no field in the ERP and no line on the P&L, but it comes out of payroll all the same.
The most expensive H2A sub-pattern is the calibration that swells until it opens a meeting. You try to fix the prompt on your own, get stuck, call the colleague who is better with the machine, another fifteen minutes. Then someone suggests pulling in the head to align on the approach. Suddenly it is four people, twenty extra minutes, three levels of seniority on the same call. Modeled cost: R$ 800 to R$ 1,200 per escalated calibration. And what was an H2A edge paraded off as an H2H edge without anyone noticing.
A2H: The Sign-Off Is the Most Expensive Edge Per Cycle
A2H is where the agent delivers and the human checks, and it is the most expensive edge per cycle of the four. The reason is simple: whoever signs off tends to be C-level, an area head or the most battle-worn senior, loaded hourly rate of R$ 480 to R$ 960. Each validation burns that hour to check an assumption, cross it against context the AI has no way of knowing, and decide whether to accept it or send it back.
When the agent gets a base assumption wrong, and it gets it wrong at a frequency you can count, the rework doubles. A2H stitches one cycle onto the next: check, ask for a fix, get a new output, check again. Two of those cycles cost what three short H2H meetings would, and nobody wrote the expense down.
| Decision type | Rework probability | Typical A2H cycles | Estimated total cost |
|---|---|---|---|
| Pricing/billing | 45-60% | 2-3 cycles | R$ 1,400 to R$ 2,800 |
| Forecast scenario | 30-45% | 2 cycles | R$ 960 to R$ 1,920 |
| Vendor contract approval | 55-70% | 3-4 cycles | R$ 1,920 to R$ 3,840 |
| Competitive analysis | 25-40% | 1-2 cycles | R$ 480 to R$ 1,920 |
| Strategic recommendation | 60-75% | 3-5 cycles | R$ 2,400 to R$ 4,800 |
The Aggregate Invoice and What Each Edge Hides
Put the four edges into a plausible distribution for a mid-market B2B company, 500 people, median AI adoption, and the map comes out like this. None of these lines shows up on its own in the forecast, in the efficiency report or in any one model's trail. Add it all into a single number and you have the right total of the wrong thing.
| Edge | % of total cost | Estimated monthly cost | Growing or stabilizing? |
|---|---|---|---|
| H2H | 45-55% | R$ 2.1 to 2.8 million | Growing: loaded hour rising + new meetings only about AI |
| A2H | 20-30% | R$ 900k to 1.4 million | Growing: more AI output demands more senior sign-off |
| H2A | 12-18% | R$ 480k to 750k | Growing fast: adoption across the whole team + iterations per person |
| A2A | 5-10% direct + remediation | R$ 180k to 380k | Growing quietly: multi-agent adoption surging |
Adding it all into one block hides two things that change the decision. H2H still leads, but because AI pushed the loaded hour up, not because of the old meeting inertia. And A2A surges alongside multi-agent adoption, charging on the human fix and not on the token. Without opening the bill edge by edge, any correction you try is a shot fired with your eyes shut. The AI Multiplier paradox unpacks how each of these distortions eats the individual gain before it crosses over to the operating margin.
Three Pragmatic Moves to Split the Bill With No New Tool
To start, you do not need to buy anything. Three moves handle the first month. The step-by-step version of the first one, with the graph drawing, the split by edge type and the consolidation into a radar, is in the edge inventory in 30 days with no new tool.
- Map the path of two or three recent decisions. Pick two or three decisions that carried weight in the last 30 days and trace the route of each one: how many humans, how many agents, in what order, how many back-and-forths. No software needed, it fits on a napkin. In three decisions, the shape of your company is already staring you in the face.
- Stick a cost on each edge using loaded payroll. Your loaded senior hour times the average time spent on the edge. For the agent, use the average inference cost (almost always low) plus the human fix it provokes (almost always chargeable). It is a rough estimate on purpose; what you want here is the order of magnitude, not the fifth decimal place.
- Re-read it every quarter, side by side with the forecast. Put each edge's number on the same dashboard where payroll and cloud spend already sit, and compare quarter to quarter. If H2H runs faster than revenue, there is a distortion eating you from the inside. If A2A blows up, the multi-agent infra is charging with nobody looking.
What Regulatory Compliance Sees and What It Does Not
AI regulation, with Brazil's PL 2338 moving through the lower house and the EU AI Act coming into force on a staggered basis from August 2026, classifies systems by risk level. It requires inventory, classification, impact assessment and traceability of every automated decision. All of that is necessary and enforceable.
Except none of it measures a coordination edge. Regulation looks at what each agent does on its own, not at what it costs to have the whole thing mesh with the humans around the decision. You can be fully buttoned up on compliance and still burn millions on badly drawn edges in the same set of books. What changes with PL 2338 and the EU AI Act in 2026 for your operation is a parallel vector, not a substitute. The direct executive read of this split (with 5 questions compliance does not answer) is in the 5 questions economic governance answers.
The Taxonomy Has Old Theoretical Roots
Splitting coordination into four edges is no slide invention. It borrows the vocabulary of two economic traditions that already had names before LLMs existed. Coordination theory, which classifies the dependencies between an organization's activities. And the theory of the firm, which explains why the company exists in the first place: to cut the cost of running each transaction on the open market. When you turn those two lenses on the human-agent mix of 2026, the old distinction comes back sharp. Coase and Williamson applied to hybrid work in 2026 ties off the other end of this thread.
Frequently Asked Questions
What are the 4 edges of human-agent coordination?
They are the four possible kinds of connection in a decision that mixes people and AI agents. H2H is human coordinating human (meeting, sign-off, the senior pulled in when the agent gets stuck). A2A is agent passing context to agent (the handoff between two pipelines). H2A is human triggering agent (the prompt refined six times until the output is usable). A2H is agent handing work back for the human to check (validation, assumption checked by hand). Each one has its own unit cost, frequency and owner in the real operation, and none of them shows up under its own name anywhere.
Why does splitting the bill by edge matter?
Because each edge fattens on a different trigger. H2H rises when you cut people and whoever stays is more senior. A2A rises when you adopt multi-agent architecture, and adoption took off over the last two years. H2A and A2H rise when the whole team starts operating with AI inside the flow. Adding the four into a single number erases exactly the information you need: where it is leaking. And without knowing where it leaks, any correction is a shot in the dark.
Which edge is the most expensive in a mid-market B2B company?
By volume, H2H dominates, in the range of 45 to 55% of the total cost of coordination in a 500-person SaaS. But per cycle, A2H is the most expensive: each senior sign-off costs R$ 480 to R$ 960 of loaded hour. The treacherous one is A2A: it looks like pennies on the token and charges on the senior engineering payroll, silently, when the handoff between agents does not close and someone has to fix it.
Is this the same as MCP, agent handoff or multi-agent orchestration?
No, and the confusion costs a fortune. MCP, agent handoff and multi-agent orchestration are technology and an architecture pattern: they answer the HOW. The four edges are an economic read of WHO PAYS for what, running across those technologies. Your company can have zero MCP in production and the four edges live in currency all the same, because the handoff between two agents is sometimes a copy-and-paste done by a senior, and a senior's copy-and-paste has a price.
How do you start measuring edge by edge with no technical instrumentation?
Three moves, no new tool. Take two or three important decisions your company closed last month and draw the path of each one on paper: how many humans, how many agents, in what order. Stick a cost on each edge, loaded senior hour for the H, average inference cost for the A. Compare the decisions and see which edge weighed in which pattern. Three decisions are enough for the order of magnitude to jump out in front of you.
The Bottom Line
The question worth asking is no longer how much it costs to coordinate. It is how much each edge of hybrid coordination costs and which of them is rising faster than your margin can take. Four bills, four triggers, four ways to stop the leak, and whoever mixes it all loses sight of the four at once.
The company that opens the four works with a scalpel. The one that looks at the single block fires in the dark and only finds out at the next close that something does not add up, without being able to point at what. In 2026 that difference starts leaking into the consolidated result. In 2027 it shows up in the price the market pays for your company.
The tool that measures this precisely is still crude; the logic that separates the edges, that one is already standing and does not need software to run. What you are left with is a choice of timing: split now, with a napkin and a loaded hour, or wait for the board to demand the bill broken out by edge on the day it is too late to improvise.