Not a mock-up. Every drawing below is plotted from this table.
MacroAgentic is your strategic co-CEO, develops and runs a plan across every agent available, tool and person you already have — and manages, to the cent, the unit economics.
FIG. 1 — Retain the Q3 renewal book resolved to an execution network. 12 tasks across 9 executors; the red spine is the critical path and the hatched tails are total float — the slack the runtime can spend before the finish date moves.
You state an outcome. It writes the work breakdown.
There is no prompt chain here. The layer decomposes the outcome into objectives, then into tasks, and attaches an acceptance test to every leaf before a single agent is woken up. It decides what needs an agent, what is a plain tool call, and what a person has to sign.
A leaf that fails its test goes back to the planner, which re-cuts the branch. That is the difference between a system with a strategy and a system with a retry loop.
FIG. 2 — Work breakdown for GOAL-0417. Assignee class is decided at plan time, not at run time; the acceptance test on each leaf is what the layer will actually check before marking it done.
The plan is the program.
The breakdown compiles to a dependency graph and the runtime executes it — wide wherever the graph allows, serial wherever it does not, with the critical path computed rather than guessed. State is checkpointed per task, so a dead agent, a rate limit or a redeploy resumes instead of restarting.
Executors are whatever you already run: Claude agents, in-house services, MCP servers, cron jobs, people. The layer does not care what is on the other end of the task, only whether it passed.
FIG. 3 — Concurrency profile for the same run. The dimensioned spans are serial pinches — stretches where the dependency graph forbids parallelism, and where adding agents buys you nothing.
Sales, marketing, engineering, training, customer success — the same machinery.
A workforce is not one team. The layer runs 5 goals at once across 5 functions, over 24 executors and 6 people, on one clock and one ledger. Sales is working the pipeline while engineering ships the billing migration and training certifies the cohort that sales is about to hire into.
Nothing about the machinery changes between them. The same decomposition, the same critical path, the same ceiling — which is the point. A function does not need its own agent stack, it needs its outcome stated.
FIG. 4 — 5 concurrent goals, one per function, on a shared clock. Red is each goal’s own critical path; the totals are the whole portfolio, not one team’s slice of it.
Cost per outcome, not cost per token.
Every token, tool call and human minute is booked against the goal that caused it. Plan and actual are integrated on the same schedule, so an overrun surfaces while there is still time to act on it — not in next month’s invoice.
This run finished $40.65 over its own plan, at 11.3% variance, and stayed inside the standing ceiling. Ceilings are enforced by the runtime: it stops work before it crosses one.
FIG. 5 — Cumulative cost, plan against actual, on the run’s own clock. The hatched wedge is variance; the dashed rule is the ceiling the runtime will not cross without a signature.
An agent gets a spend limit, like anyone else on the payroll.
Each executor carries a standing limit and a blast radius. Work that would breach either is escalated for a signature instead of attempted, and the refusal is recorded alongside the spend. On this run voice-01 came within a few dollars of its cap — visible here while the run was still going, rather than discovered afterwards.
FIG. 6 — Spend against standing authority, per executor, for GOAL-0417.
Every run closes with a reconciliation.
The same table that scheduled the work settles it. Which executor earned its place, which one ran hot, and what the outcome cost in total — reconciled per run, not estimated per month.
| Executor | Class | Tasks | Minutes | Plan | Actual | Var. |
|---|---|---|---|---|---|---|
| voice-01 | AGENT | 1 | 34 | $124.00 | $149.80 | +25.80 |
| analyst-01 | AGENT | 3 | 43 | $60.00 | $70.30 | +10.30 |
| rev-ops | HUMAN | 1 | 26 | $52.00 | $52.00 | — |
| vp-success | HUMAN | 1 | 20 | $40.00 | $40.00 | — |
| support-04 | AGENT | 1 | 28 | $33.50 | $36.20 | +2.70 |
| writer-03 | AGENT | 1 | 18 | $29.60 | $31.10 | +1.50 |
| ops-02 | AGENT | 2 | 25 | $16.10 | $16.45 | +0.35 |
| postmark.mcp | TOOL | 1 | 6 | $3.10 | $3.10 | — |
| warehouse.mcp | TOOL | 1 | 8 | $1.20 | $1.20 | — |
| Total — GOAL-0417 | $359.50 | $400.15 | +40.65 | |||
Every run teaches the next one.
Plan, dispatch, measure, re-rank, plan again. Outcomes and costs go back into the registry, so the next composition ranks candidates on what they actually delivered rather than what they advertised. An agent that runs hot gets picked less; one that quietly beats its price gets picked more.
Across six passes at the same goal, cost per outcome fell 18.1% and then flattened — the loop converges rather than chasing its tail.
FIG. 7 — The control loop, and what it bought. Each pass re-ranks against measured cost and outcome; the curve flattens when further swaps cost more than they save.
Applies to the whole sheet unless a detail says otherwise.
Bring one outcome you currently chase across four tools and three people. We plan it, run it and hand you the sheet.
Get started