Your Job Is the System, Not the Code

The short version: stop writing the code and build the system that grades the agents who write it. Between 25 September and 12 October 2026, Claude Code agent swarms built two products while I typed short prompts. Eight rules came out of it.

Rightmark helps companies license private data to AI developers. Dylan is an AI SDR that books B2B meetings, billed when held. I ran both builds at once on one Mac.

1. The build stopped being the slow part

Across 18 days and 20 sessions I typed 223 prompts, and the median one was 52 characters. This is two first-hand builds, not a benchmark. Ten of those prompts asked whether it was all done and 15 asked what is left. The typing is not the slow part any more. Reading the evidence and deciding what to accept is, so the job is now: ask for status, decide, and send work back.

2. Stop being the builder

The Rightmark swarm rules give the lead session one job description: it “plans, delegates, merges through one integrator, and reports. It does not write application code.” The Dylan lead prompt puts the same thing in four sentences: “You lead. Specialists build. Independent agents check. Scripts hold the plan.” Write that rule into the session prompt before the first unit: once the session in charge starts coding, nobody holds the whole plan.

3. Write work as contracts someone else can check

Each Rightmark unit is a card: dependencies, the files it owns and reads, acceptance criteria, evidence, review. Shared files belong only to integrator units. Decisions run on defaults: the swarm uses the default until I decide, and open items wait for the next gate. The decisions file holds 27 rows and 24 dated answers. On Dylan, a planning pilot lane checked 58 facts: 47 confirmed, 11 corrected, 0 refuted. Pilot one lane before you commit the plan, and make the pilot report a count: facts confirmed, facts corrected, facts refuted.

4. Never let the builder grade its own work

“The agent that built a unit never grades it.” A fresh verifier reruns every command, and a second agent tries to disprove every finding. Of 3,021 agent starts in 85 Rightmark workflow runs from 6 to 12 October, 1,353 were checkers and 337 were builders: four checkers for every builder.

Checkers need a stop rule too. On 6 October the scaffold unit burned three hours and came back blocked with zero P0 or P1 findings, because the verifier kept inventing new bypasses. Every other unit waited. Now only P0 and P1 gaps block.

5. Taste does not delegate

On 8 October rightmark.ai went live, and the same day I sent the homepage back: “its not the website we initially design, that is not the main page i prepared at the beggining.” An agent’s style rule had silently rewritten my approved hero copy. Now the design I approved on 4 October is the reference for every page in it, and only a hard rule may force a difference, with each one listed. All 15 parity units are merged. Agents can hold a style. Only I could say which one was mine, so the approved design goes into the project rules as binding and parity runs in the full gate before merge.

6. Automate the clicks, keep the signatures

On Rightmark, a short list waits for my explicit OK: production deploys, DNS, purchases, emails to real people, pushes, publishing, accepting terms, and guardrail changes. I gave the production OK on 8 October. The Scout terms are a draft for counsel, and no health data is taken until a HIPAA program and vendor agreements are in place. Those are signatures, not clicks.

For Dylan, on 9 October I typed “Please i dont wnat to do anything manualy use chrome and do all of it”, and the next day “i will handle leads, payments manualy for now …”. Live payments stay off until counsel approves the Terms, and the client approves the plan, the first emails of every new angle and any new spend. I refused manual clicks and kept the money.

7. Test where the customer is

On Rightmark, a watcher script did the shipping. Every 10 minutes it deployed a clean export of main when main had moved, smoke-tested every sitemap URL, rolled back on a failure, and stopped after three failed deploys in a row. From 8 to 12 October, 201 production deploys passed their smoke test. Two more attempts failed and both were caught: one never reached production, and one broke sign-in with a 500 and was rolled back on its own. A sibling session also reported that 500, at 10:30 UTC on 9 October, and noted that the smoke list did not include sign-in.

On Dylan, only a production end-to-end found every waitlist step after the first answering 503: the storage layer served a weak etag the update check never matched, and local tests used a file store. Passing tests describe your laptop. The customer lives in production.

8. Some days poorly: the machine and the clock

On 11 October the parallel slots across Rightmark runs reached 11, the Mac’s load hit 460 and gates timed out. The same day a test leak filled the disk and 12.6 GB had to be cleared. The clock is a budget too. One build prompt read: “Pacing: the 5-hour usage window is at 72 percent and resets about 18:30 UTC.” On 7 October, and on 12 October up to the morning snapshot, the Rightmark sessions ran with zero prompts from me. Budget the machine the way you budget people: about 5 in flight, an alert below 12 GiB of free disk, and the day planned around the usage window that resets.

If you own data, Rightmark gives an indicative range in about a minute at rightmark.ai/appraise, with no files and no access. If you own code, Repo Check values a Git repo on your own machine and never uploads the code. If you need B2B meetings, join the Dylan waitlist; he bills per held meeting and has no results to show yet.

Connected reading

Continue through AI Operators Need SOPs, Not Prompts, Human Review Is Becoming the Premium SKU, Distribution-First Products and the AI agents for operators hub. The full build record is The Conductor Never Codes on Dive.

Sources and methodology

Dated 12 October 2026. Counts come from session transcripts, workflow journals and result files, git logs and the Rightmark deploy log, frozen at the 12 October snapshot while Rightmark sessions were still running. Main-session tokens were not measured, and no dollar figures are given. This AI-assisted first-person draft follows a no-fabricated-benchmarks rule.