Darroman — Managing AI Agents: How One Founder Ships 25 PRs a Day, Solo
TL;DR
A third-party write-up of Ryan Carson’s operating system at Untangle, published a month before the How I AI episode that covers the same ground. It earns its own page because it is the numerate version: where the episode gives reasoning and on-screen demonstration, Darroman extracts the cadences, the per-run costs, and the target unit economics. Carson ships 22–25 PRs a day, sometimes 40, alone.
1. Cloud over local — the same hard line, with the mechanism named. “If you are working locally, I honestly think you are a caveman.” The argument is about collisions: cloud agents give an isolated virtual machine per session, so parallel work scales horizontally. Carson runs 5–10 cloud agents simultaneously — “at least 5, often 10.” The article makes the structural point explicitly: the PR volume comes from many ordinary agents in parallel, not from one very good agent. That is the same primitive Geng & Neubig isolate and test as isolated workspaces in 2026-03-23-geng-neubig-caid-asynchronous-software-engineering-agents.
2. Decision load explodes before output does — the article’s best contribution. Where a developer might make 2–3 high-stakes decisions a day, Carson faces 10–20 before lunch. The countermeasures are all attention-rationing devices rather than tooling:
- Pin priority threads — only 3 of ~25 daily PRs get focused attention
- A 25-minute check cadence — bounded review intervals, not continuous supervision
- A paper to-do list for the week’s priorities
- Roughly half his work from an iPhone browser
- Agents never hold production credentials; sensitive actions are approved by hand
The last of these is a least-privilege posture arrived at by a practitioner from operational necessity, matching what Singapore’s regulator prescribes from first principles in 2026-01-22-imda-model-ai-governance-framework-for-agentic-ai — its third pillar being technical controls including access limited to whitelisted services.
3. Three standing automations (“sleeping automations”).
| Automation | Cadence | What it does | Cost |
|---|---|---|---|
| End-to-end signup test | 3× weekly | Agent completes the full signup workflow, records video, identifies bugs, files fixes | ~$60/run |
| Production watchdog | Daily, 9am | Reviews previous 24h of customer activity; summarises insights and anomalies with session links | — |
| Self-improvement loop | Daily | Grades Grace (Untangle’s AI paralegal) conversations against a quality rubric; substandard interactions trigger child agents that open fix PRs | ~$5/session; ~3 fixes ship/day |
4. Token economics, and the number that matters. Current spend $20,000/month, which Carson calls unsustainable; the estimated sustainable level is ~$5,000/month per engineer-equivalent. The route from one to the other is model routing — frontier models for complex reasoning, fine-tuned coding models for repetitive work — enforced structurally through parent–child agent patterns that keep premium models off routine tasks. His business-model prediction: $200/month all-you-can-eat plans will not survive, and he favours independent platforms (Devin, Factory, Amp) that optimise model economics for production use rather than for a subscription price point.
5. An implementation ladder. Run one cloud agent session → run three simultaneously → build one automation (the production watchdog is named as the highest-signal entry point) → add a quality-grading loop on your core product surface. This is the most actionable on-ramp in the practitioner material, and it deliberately puts observability before autonomy.
Core thesis: managing AI agents is the emerging founder skill set; technical depth now develops through delegation and judgement rather than through typing lines of code.
The third automation is a shipped instance of the self-improving-agent pattern documented at YC — with the important difference that Carson’s loop is bounded to a quality rubric on one product surface rather than aimed at the product itself, which is the boundary he defends explicitly in the How I AI episode.
Dynamic-capabilities reading
digital-transforming/redesigning-internal-structures— the three standing automations replace what would otherwise be QA, customer success, and quality-assurance functions; the article is explicit that agents substitute for hires.digital-seizing/balancing-digital-portfolios— model routing and parent–child patterns are literally portfolio balancing at the inference layer: expensive capability reserved for where it changes the outcome.strategic-renewal/business-model— both Untangle’s own model (one person, venture-scale SaaS) and Carson’s prediction about agent-vendor pricing are business-model claims.
Linked entities and concepts
- Entities: Ryan Carson, Untangle, Devin, Cognition
- Concepts: agent-fleet-management, agentic-pull-requests, agent-oversight-and-delegation, agentic-engineering
- Dangling (single-source mention, deferred): Florian Darroman
Scope and reliability
Secondary reporting on a single self-reporting practitioner. Darroman is summarising Carson’s public accounts, not measuring anything; every figure traces back to Carson. Where this page and the episode disagree — “5–10 agents” here versus “10, 15 threads” and “15 agents” there, “22–25 PRs” here versus “40 PRs” there — the discrepancy is a month of drift plus loose counting of threads versus concurrently running agents, not a contradiction worth resolving. Use the ranges, not the point estimates.