Operating notes · Note 10
September 7, 2026 · 7 min read
Agentic practice
Agentic

The economics of
agent-operated software

The published bill for the pivot is $672 in Cursor, over 27 days, against a $10,000 engagement that did not finish the system. Zero full-time developers is not the same sentence as free.

↓ Figures as published, and only those
01
Two bills

$10,000 did not finish it.
$672 changed its shape.

Built From the Rubble is the source for every dollar in this note. The original engagement was framed as a $10,000 project that could not be finished. The team was an outsourced group in the Middle East. They left a real stack — ASP.NET Core 9, PostgreSQL, RabbitMQ, Vue.js — and a data model that was live on April 2, 2025, at 36 tables. They did not leave a system a non-coder could keep alive. Late in 2025 the team was offline, in the essay’s account because of conflict and war displacement, with production still tied to that zone.

The pivot away from that shape took 27 days and $672 in Cursor usage. The essay calls that the published cost of transforming the microservices system into an agent-ready monolith: architectural planning, the code transformation, the documentation, and the protocol layer that lets Qwen operate the system day to day. The rebuild shipped with zero critical defects in production. The method behind the defect claim is not published. I will not supply a test budget or a bug count to decorate it.

These are different bills. $10,000 is the engagement that produced the prior system and did not finish it. $672 is Cursor spend on the pivot, not a replacement invoice for that engagement, and not a statement that the earlier work had no value. The data model was real. The agents could not operate it until the shape changed.

$10k
Original engagement, unfinished
$672
Cursor spend for the 27-day pivot
27d
Microservices to the agent-ready monolith
0
Full-time developers in the operating claim
02
The weeks

Four published weeks.
Not a second invoice.

The essay also publishes four weekly buckets. It says, in the cost card, that those phase totals are the published weekly figures and that they are not a second invoice. This note will not add them up and call the sum a correction to $672. It will not explain a gap between the buckets and the total. The gap is not given a cause in the essay, so it does not get one here.

Published week
What the essay says the money was for
Feb 16–22 · $207
Peak intensity. Architectural reckoning. $44 of it on February 21 alone. Planning the move from RabbitMQ to Hangfire.
Feb 23–Mar 1 · $163
Execution. RabbitMQ out, Hangfire in. The golden flow collapsed into one readable path.
Mar 2–8 · $118
Hardening. The agent protocol layer. The essay names AGENTS.md, QWEN_PROTOCOL.md, HANDOVER_TO_QWEN.md, and CHANGELOG_AI.md.
Mar 9–14 · $96
Stabilisation. Handed to Qwen Coder. Spend dropped. Bede entered staging.

The $118 week is where a reader will want the protocol files quoted. They can be named. Three of them could not be quoted, because the bodies were not available. The economics of that week are still the published $118 and the claim that the system became legible to machines. Legibility, in the files that do exist, is the AGENTS.md router and the backend locks. I will not price a document I cannot open.

What the money bought, stated only at that resolution, is a redesign: fat controllers, named Hangfire jobs, one deploy command, and a doc layer the next operator could read. It bought a drop in the need for the heaviest sessions. It did not buy a published revenue line.

03
Steady state

After handover,
the published run rate is small and specific.

Expected daily Cursor and Qwen spend after the pivot, for routine operations, was about $15. The scorecard’s separate line is that Qwen operates on about $5–15 a day once the docs exist. I am citing both sentences. I am not adding them into a $20–30 day, because the essay does not say the $15 excludes Qwen. The first sentence says “Cursor and Qwen.” The second describes Qwen’s band once documentation is in place. A reader can hold both. A spreadsheet that stacks them would be my invention.

The 95.2% cache hit rate is the published reason the documentation changes the run rate. Files written once are read across sessions instead of being regenerated. The rubble essay’s sixth rule is the same idea in operating language: treat docs as the real codebase, because at that hit rate they are read hundreds of times. An hour on a runbook saves days of re-orientation. Orientation itself moved from fifteen to twenty minutes to about two. That is a cost, measured in session time, even when it is not on the Cursor invoice.

~$15
Expected daily Cursor and Qwen, routine operations
$5–15
Qwen per day, once the docs exist
95.2%
Cache hit rate published with that claim
~2m
Orientation after AGENTS.md

Zero full-time developers is the headcount claim next to these rates. It is not a statement that the marginal cost is zero. The daily band is the published marginal cost of the agents. Production deploy and rollback still require a human approval in GitHub Actions. Never push directly to main. The root AGENTS.md chain is feature branch, tests pass, pull request, merge, auto staging deploy. That person’s time is not given a rate. The Complete Story describes the rebuild period as no salary and no team. I will not convert “no salary” into an implied wage.

04
The comparison

$80,000–150,000
is a case-study comparison.

The same cost card says a five-person team for that scope sits at $80,000–150,000. The essay presents it as the comparison in the case study, not as a quote Steller received and declined, and not as money spent. This note keeps that label. It is a counterfactual printed beside $672. It is not an accounts-payable line, and it is not a salary benchmark I can apply to a hiring plan. No payroll, no bid, and no team-of-five proposal is published beyond that range.

The useful reading is the one the card already forces. The pivot’s published tool spend is hundreds of dollars. A human team at the case study’s range is a different kind of number. Whether that team would have finished the microservices system is not a fact the essay establishes. The fact it establishes is that the agents could not operate the old shape. The Complete Story’s sentence is the economic one, even without a price. The AI did not rescue the platform. The redesign made it rescuable by AI.

Bede’s integration is the published commercial event next to the $96 week: the Zain Sudan digital wallet entered staging, and later the essays describe an active API integration and a first successful partner transaction in under sixty seconds. Contract value, take rate, and margin are not published. Leaving them out is not modesty. It is the edge of the ledger.

05
Where it breaks

The economics hold
inside the published constraint.

The cost structure breaks where the architecture essay says the pattern breaks. A team past a handful of engineers reintroduces the merge conflicts fat controllers create, and it reintroduces salaries this ledger does not contain. Independent scaling — the essay’s example is a Bamboo path that needs many times the resources of catalog sync — is a different infrastructure bill than docker compose up -d and a Postgres-backed Hangfire. A domain of dozens of flows stops being readable, and unreadability puts the fifteen-to-twenty-minute orientation cost back, which puts token spend back. Steller’s published position is one golden flow, not forty. This note does not claim the cheap steady state survives an exit the product has not taken.

The economics also break if the docs are missing. The $5–15 Qwen band is conditioned on docs existing. The three protocol files named in the $118 week are not quotable here. The files that are quotable — the 63-line router and the 121-line backend locks — are the ones the daily rate is allowed to depend on. CORS, worker count, no-retry plus refund, the Bamboo Idempotency-Key, JWT_KEY from the environment, strict catalog validation, and the business-claims middleware are cheap to obey and expensive to rediscover in a session. A 95.2% cache hit rate does not help a rule that was never written down.

What I will not add, because it was not published: revenue, gross margin, Bede’s price, a forecast, a customer count, or a claim that $672 is the all-in cost of the company. Hosting, vendors, and the founder’s unpaid time sit outside the Cursor number. Zero full-time developers means the published headcount of that role is zero. The bill that remains is the one the essay prints, the human approval on production, and the shape of the system that makes a small daily model spend enough.

Next in the series
Note 01 · Operating notes
What Belongs in AGENTS.md
Router, hard rules, and the locks an agent is not allowed to relax.
Read the essay →