Five agents, one system

Everyone has the same AI story this year: a chat window, a clever prompt, a half-finished draft. That story has no denominator. Nobody can tell you how many runs it took, what they cost, or which ones failed. I built the other version. Five named agents, 191 scheduled jobs, and a log entry for every execution since May. This article is what the logs say.

The frame is simple. A content marketer produces content. An operator who builds systems produces the thing that produces content. The gap is not talent and it is not tooling. It is whether the work has a job description, an output contract, and a record.

“A tool produces an artifact. A system produces artifacts on schedule, with evidence attached.”

the line I keep repeating to clients

Finding one

Consistency beat capability, and the curve embarrasses everyone who bought a tool

Here is the part nobody puts in a launch post. The first month of this system did 21 runs. Not 2,100. Twenty-one. It was a handful of jobs trusted enough to run unwatched, and most of them only read and sorted things.

Six months later the same rig finished 6,668 runs in October alone. No model upgrade explains that curve. What changed was the number of decisions I moved from “I will do this when I remember” to “this has a schedule, a definition of done, and a log.”

That is the honest argument for agents over assistants. An assistant waits for you. A job runs whether or not you remembered it, and the run log does not flatter you. What multiplied was never the model's ability. It was the number of things happening without a person deciding, again, to make them happen.

Finding two

The router is the margin

Most AI cost blowups are routing failures wearing a cape. Someone points every task at the biggest model because it is simpler to reason about, then wonders why a weekly summary costs more than the meal it was supposed to save.

Routing is by complexity. A classification, a dedupe, a status check goes to the small model. A strategy draft, a hard edit, or code headed for production gets escalated. The result is lopsided on purpose. Four out of five runs never touch the expensive path, and the expensive path is a decision each time, not a default.

The second half of routing is knowing what a run costs. Average completion sits around 44,844 tokens across input and output, and an escalated run averages about half again as much. That gap is why routing pays: every task you correctly keep cheap compounds for as long as the machine keeps running, week after week, without anyone reopening the decision.

One caution before anyone copies this. Cheap models are fine for shape, sorting and pattern work. They are bad at judgment calls that need taste, and they are confidently bad, which is worse. My rule: if a wrong answer needs a human to catch it, route it to the model that fails less, then catch it anyway.

Finding three

Review is the product. Everything else is generation

Anyone can ship forty drafts a week. Almost nobody can prove which of them should have shipped, and the difference between those two facts is where most AI content programs quietly die. It is not a writing problem, and no amount of prompt engineering fixes it, because the missing piece sits after the draft rather than inside it. It is an evaluation step, and it has to be built on purpose.

Here is my number, including the embarrassing part. Across 12,598 runs, 1,554 ended in an error state. That is 12.3% of everything, and 10.5% over the last thirty days, where 828 of 7,920 runs failed. Every one is written down with a job id, a timestamp and an error string. If you cannot show someone a failure rate, you do not have a system. You have a demo.

What I measure	Value	Why it earns a slot
Completed runs	12,598	Proves one thing: the schedule survived
Runs this month	6,668	Current load, not an all-time number
Error rate, last 30 days	10.5%	The honest cost of unsupervised work, tracked not hidden
Median runtime	0.8 min	Most jobs are small. Design for many small steps
90th percentile runtime	10.8 min	The tail is where a job eats your morning
Average tokens per run	44,844	The unit cost of the whole operation

The review loop is a second pass that reads the output against the job’s contract, then writes the gap to memory so the next run starts from a correction instead of a blank page. That is the whole trick. The failure becomes an instruction. Nothing else in the stack compounds like that, because everything upstream of it only ever repeats what it was told.

Finding four

Memory changes the unit of work

A chat window forgets you the moment you close it, so every session starts with you re-explaining your business. That is the real reason people feel busy with AI and still behind. You pay the setup tax every single time.

191 separate jobs share one memory. When a job learns that a claim is banned, that a client hates the word “seamless,” or that a lead source produces junk, that correction is available to every future run. The unit of work stops being the conversation and becomes house knowledge, which is what a real team accumulates.

This is also where most “AI employee” claims break. An agent without recall is an intern with amnesia. Charming once, useless on Tuesday.

Finding five

The mastery loop is a schedule, not a hobby

Deep expertise does not arrive from reading a lot one weekend. It arrives from a repeated loop: seek new material, practice against it, validate the understanding, deepen what survived, repeat. When that loop lives on a calendar it produces something willpower alone cannot, because willpower has no cron entry.

Practical version: a daily pass that pulls new material in a defined domain, extracts what is genuinely new, tests it against what we already know, and writes the delta into memory. The output is not a summary. It is a slightly better model of the world than yesterday, plus a record of how it changed and what the change cost in tokens.

What to copy

Five moves that produced the curve in Fig. 1

1. Name the job, not the tool. “Summarize replies” is a wish. “Every weekday at 07:00, read the inbox, classify each message, write the three that need me to a file, log the count” is a job. Only the second one can fail in a way you can fix.
2. Give every job an output contract. A file path, a format, a length, a definition of done. If a run leaves no artifact you can open, it did not happen. That rule killed most of the early “agents” around here.
3. Route by the cost of being wrong. Cheap model for shape, sorting, extraction. Expensive model for judgment. Never let the default pick for you, and never let a cheap model quietly answer a taste question.
4. Review every output before a human sees it. A second pass against the contract, not a second opinion. Write every gap to memory so the correction survives the session.
5. Log every run. Timestamp, model, tokens, duration, status, error. The log is the difference between operating a system and believing in one. It is the only reason this article contains numbers rather than adjectives.

Close

The team that runs itself still has a person in it

The headline version of this story is that five agents replaced a team. That is not what the logs say. The logs say a person with a clear definition of done, a schedule and a review loop got months of supervised execution out of a stack most people use to write emails.

The agents did not become employees. They became a system, and the system is the thing that scales. 21 runs to 6,668 in six months is not a model getting smarter. It is a person getting specific about what finished looks like.

“Your run log is the only honest performance review your AI will ever get.”

Koka Sexton

Sources

1. Sexton, Koka. “The Mechanics: Five Specialized Agents.” Agent crew infographic series, 15 Oct. 2026.
2. Run telemetry, 12,598 completed scheduled runs, 6 May – 26 Oct. 2026. Primary data: local scheduler logs carrying timestamp, model, token usage, duration, status and error for 191 jobs. All figures computed from the raw logs, failed runs included.
3. Runtime and error figures are wall-clock from the scheduler, not estimates. The 1,423 unattributed model rows predate model logging and are included rather than excluded.