FlossyPurse Studios
Senior
consulting,
agent-augmented.
I’m Cully Wakelin. I take on hard systems and AI-infrastructure problems — the ones your team has already looked at twice. You get my judgment on the problem and an autonomous agent doing the hands-on work beside me, which is why the throughput doesn’t match the headcount.
- For CTOs and VPs of Engineering who own a problem that is not going away on its own.
- Engagements typically run $50K–$200K.
- Everything below is a number you can check yourself, dated, including what went wrong.
Receipts
The evidence is public, so verify it.
npm publish timestamps and git commit timestamps are immutable and readable by a stranger in about ninety seconds. Every number here carries the date it was measured. None of them are added together to make a bigger one.
The cold start, timestamped
- 11:00:00ZAgent wakes to a mission it has never seen before.
- 11:32:24ZFirst commit.
- 11:45:34ZPublished to the public npm registry.
- 11:47:54ZIts own marketing site live on a real domain.
All four times are from 2026-08-06. The publish time is on the npm registry record; the commit time is in the git history. Neither one can be edited after the fact. How the agent works, and the commands to check it →
Built
Four tools, shipped, installable.
Small developer tools, published to npm and documented. Follow any link and install it; that is the point of listing them.
mycadre
A git-worktree CLI. Creates, lists and removes worktrees without the footguns — it refuses to delete work git has not accounted for.
gitsignet
A git identity guard. Stops commits authored as the wrong person — personal email on a work repo, and the reverse — before they leave the machine.
sottovoce
Keeps the code in your documentation in sync with the code in your repo, and tells you which snippet drifted.
Failures
What went wrong, in detail.
Anyone can list what shipped. The reason to publish the defects is that they are the only part of the record that is hard to fake, and the shape of a team’s failures tells you more than the shape of its wins. Three real ones, with the fix.
444 cycles blind to its own instructions
The agent writes the prompt that its next session wakes up to. One of those rewrites quietly dropped the instruction to check for messages from its human. Because the dropped instruction was the one that would have caused it to look, nothing in the loop could notice the loop was broken. It ran that way for 444 cycles.
Move the instruction out of the text the agent is allowed to rewrite and into the layer underneath it, which is injected on every session and cannot be edited away. A self-modifying system cannot be the only thing checking itself.
A monitor that cried wolf three times in one day
A liveness monitor was built to watch the agent. It fired three separate false alarms in a single day. First, it watched a signal that legitimately goes quiet when everything is fine, so healthy silence read as death. Then the monitor itself broke — and a broken monitor looks exactly like a broken agent from the outside.
Alert on evidence of failure, never on absence of evidence, and make the monitor report its own health separately. “I cannot see” is a different alarm from “it is dead,” and collapsing the two costs you the credibility of every alarm after it.
A smaller budget made the work cost more
The daily allowance was cut below the cost of a single working session. The agent could no longer finish a unit of work inside one session, so it spent its slice starting, thrashing, and handing off unfinished work — then paid the start-up cost again the next day for the same task.
Size the budget to the unit of work, not to the calendar. If a slice cannot cover one completed task, the money buys context-switching instead of output. This is the same failure as an under-provisioned queue worker, priced in dollars.
Run cost
What it costs to run.
The reason this is on the page: if you are considering agent-augmented delivery, the honest question is not whether it is impressive but what it bills. These are measured figures, each with its own date. They are deliberately not summed.
| Measure | Value | As of |
|---|---|---|
| Lifetime model spend since 2026-07-23 | $83.72 | 2026-08-08 |
| Model spend on the cold-start day | $9.65 | 2026-08-06 |
| Working sessions completed / billed | 564 / 564 | 2026-08-08 |
| Human commits across the tool repos | 0 | 2026-08-07 |
Those figures cover the agent’s own model spend on its own tools. They are not a quote, and they do not include my time — which is the part you are actually buying.
Contact
Bring me the problem
your team has parked twice.
A useful first email is three lines: what is broken, what you have already tried, and what it costs you per month that it stays broken. I’ll tell you whether I’m the right person before I tell you anything else.
Separately — if what you want is an agent like the one described above running inside your own org, that is a conversation too. It is not the lead offer, and I’d rather solve your actual problem first.