Most Valuable Skill of 2026: Managing AI Agents

Ryan Carson, sole employee of AI divorce startup Untangle after raising a seed round, argues that managing teams of AI agents will be the most valuable skill for founders, solopreneurs, and students in 2026. Untangle serves family law firms and Carson projects its revenue will quadruple in a single month. His prior company Treehouse grew to 110 employees and taught a million people to code, giving him a direct basis for comparing traditional team management with AI agent management.

Carson runs five to ten cloud-based AI agents simultaneously, typically shipping twenty-two to twenty-five pull requests per day and sometimes reaching forty. He works across an eight-screen layout on a 52-inch Dell monitor and completes roughly fifty percent of his work from an iPhone, which he treats as a practical requirement for staying available to agents continuously. Agent status updates arrive through Slack and most text input goes through Whisper Flow via a dedicated button.

The shift to cloud-based rather than local development is central to his output. Running multiple agents locally requires multiple code copies or complex git worktree management, creating collision risk and limiting parallelism. Cloud harnesses like Devin from Cognition spin up virtual machines on demand and eliminate those problems. Carson estimates that working locally produces roughly ten times less output. He chose Devin after concluding that when it launched approximately two to two and a half years ago the underlying models were not strong enough, but that as of 2026 the models are sufficient. Other harnesses he identifies include Codex from OpenAI, Claude Code from Anthropic, Amp, Cursor, and Factory.

The nature of work shifts heavily toward high-stakes decision-making. Carson estimates he makes ten to twenty high-stakes decisions before lunch, compared to two or three per day in a traditional workflow. He pins the most important agent threads, separates them from smaller bug fixes, and checks high-priority threads approximately every twenty-five minutes. He argues that technical understanding becomes more important rather than less in this model because directing agents effectively requires knowing concepts like Postgres, production versus development environments, and database migrations. On security, he does not give agents production write keys and instead manually copies them from 1Password, on the grounds that agents will eventually do something harmful if given persistent production access.

Cost management is handled through a parent-child agent architecture. A parent model using Fable manages orchestration and spins up child Devin sessions running a cheaper fine-tuned coding model called SWE 1.7, which Cognition built specifically for coding tasks at lower cost than frontier models like Opus 4 or GPT-5. Carson spent approximately twenty thousand dollars in tokens last month, which he described as unsustainable, and estimates a viable AI engineering cost per employee is closer to five thousand dollars per month. Three specific automations illustrate the system. An automated end-to-end signup test runs three times per week, costs roughly sixty dollars per run, and triggers a triage session that spins up a new agent to fix any failures found, with Devin recording and annotating browser video and reviewing its own recording in an agentic loop. A production watchdog runs daily at 9am, scans all database events for paid customers, and surfaces a summary in an admin dashboard with direct links to live UI instances. A self-improvement loop grades daily chats from Untangle's AI paralegal agent Grace against a rubric, and any chat scoring below threshold spawns a child session to fix the issue and open a pull request, producing approximately three shipped fixes per day at a cost of roughly fifteen dollars per day total.

Despite this level of automation, Carson maintains there is currently no substitute for a founder personally using their own product to catch UX and quality issues, and that even capable AI models still lack obvious common-sense judgment about acceptable product experiences. High-stakes decisions cannot yet be fully delegated.

On tooling strategy, Carson warns against building entirely on a frontier lab stack like Claude Code or Codex because doing so creates vendor lock-in. He believes Claude Code and Codex are subsidizing token costs in a way that is not sustainable. Independent agent labs such as Devin, Amp, Factory, and Cursor are incentivized to optimize for cost through model routing while still allowing use of frontier models internally. His recommendation is that founders who have reached product-market fit and are beginning to hire should move to an independent agent lab rather than a frontier lab walled garden, while solo founders building early-stage software can likely manage on a two-hundred-dollar-per-month frontier lab subscription before needing a full software factory setup. He closes with the caveat that the tools and best practices he describes will change every three to six months.