NanoClaw and the Recurring Agent

Six weeks after the first commit, NanoClaw has a $12 million seed round and the attention of every CTO who has been quietly cron-ing a Claude script in production.

NanoCo, the Tel Aviv company behind the open-source agent harness, announced the round on May 20, 2026. It was led by Valley Capital Partners, with participation from Docker, Vercel, Monday.com, Slow Ventures, Clutch Capital, Factorial Capital, and Hugging Face CEO Clem Delangue as an angel. According to Fortune, the brothers behind NanoCo, Gavriel and Lazer Cohen, turned down a roughly $20 million acquisition offer to take the seed instead.

The pitch from the lead investor is the kind of line that lands on a deck slide. Steve O'Hara of Valley Capital Partners: “AI is shifting from a tool that answers questions to become an assistant that can get work done for every business worker.” Gavriel Cohen, talking to SiliconANGLE the same day: “They're asking how to roll it out to their teams. They've figured out where the value actually lives: an agent has to be able to work inside the most sensitive parts of a business.”

What NanoClaw Actually Is

NanoClaw is a sandboxed alternative to running Claude Code or similar harnesses directly on a developer machine. The agent lives inside a Docker container with no ambient credentials; secrets are injected at runtime by a separate gateway, and sensitive actions require a human approval ping in Slack or WhatsApp. The GitHub repo sits at roughly 29,500 stars as of May 27, 2026, with installable channels for WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, Linear, GitHub, WeChat, and email.

What makes the README worth re-reading is one line that hides between the channel list and the security model:

“Scheduled tasks - recurring jobs that run Claude and can message you back.”

And, three sections later: “each agent group has its own CLAUDE.md, its own memory, its own container, and only the mounts you allow.” That pairing is the actual product. A long-lived agent with a name, an inbox, a persistent workspace, and an instruction file that follows it across runs. The schedule is not the agent. The agent is the agent.

The Surface Nobody Was Building Against

For two years the agent ecosystem shipped in two shapes. The interactive editor: Claude Code, Cursor, Windsurf. The autonomous one-shot: Codex, Devin, Replit Agent. Both are reactive. You start the agent, it runs to completion, it stops. There is no Tuesday morning at 9am where the agent does a thing because Tuesday morning at 9am is when that thing happens.

A cron job that wraps a CLI is not the same thing. A cron job is a script that calls a model and forgets it. What NanoClaw shipped is closer to a coworker: a named agent that wakes up on a schedule, reads the same CLAUDE.md every time, writes to the same memory store, and pings you on the same channel.

In an interactive editor, a system prompt configures a session. In an autonomous one-shot, it configures a single task. In a recurring agent it configures an identity that wakes up over and over with the same character, the same priors, the same opinions, and the same memory of how last Tuesday went. That is not a session prompt. That is a persona. The best constructs are opinionated; recurring agents are where that argument stops being aesthetic and starts being load-bearing. Run an opinionless agent fifty times and you get fifty adequate summaries and no point of view. The persona is what compounds.

Three Recurring Agents Worth Writing

Each of these is a single named agent with one construct attached and a schedule. Each is more useful on the hundredth run than the first, because the construct holds the standards and the memory holds the history.

Nightly code review. Every weeknight at 11pm, an agent named Margot reads the day's merged pull requests, drafts review comments in the voice of a reviewer with opinions about test boundaries and naming, and posts them to a private Slack channel for the engineering manager to skim with coffee. The construct is a Senior Code Reviewer; the memory tracks which authors have already been told the same thing twice.

Weekly tabletop drill. Every Friday at 3pm, Hollis runs a plausible outage scenario, walks the on-call engineer through triage, and writes up the gaps the drill exposed. The construct is an Incident Commander that knows the company's services and refuses to award credit for vague answers. The memory tracks which playbooks the team keeps failing.

Sunday brand audit. Every Sunday at 6pm, Coda reads the live landing page, the three most recent posts, and the pinned tweet. It flags copy that drifted off-brand, headlines that buried the lead, and CTAs that contradict the positioning. The construct is a Brand Strategist; the memory remembers which critiques were taken and which were ignored.

None of those agents are exotic. The exotic part is that they have names, schedules, and continuity. NanoClaw is the first harness to make that ergonomic without writing your own orchestrator.

Why the Persona Has to Be Portable

The harness layer is in flux. NanoClaw is the loudest entrant this month, but Anthropic ships an Agent SDK with scheduling primitives, OpenAI has equivalents in flight, and within two quarters the recurring-agent surface will be competitive across every major provider. The harnesses will compete on isolation, channels, observability, and cost.

What they will not compete on is the persona. The persona is upstream. The harness can be replaced quarterly without breaking the agent; the persona cannot be replaced without replacing the agent. This is the format-war thesis from The Cambrian Explosion of Instruction Files applied to scheduled work. CLAUDE.md, .cursorrules, AGENTS.md, SOUL.md, and whatever NanoClaw standardizes on, all read roughly the same Markdown. A construct written for a NanoClaw agent today is the same construct that drops into Claude Code tomorrow as a project-level CLAUDE.md, into Cursor as a .cursorrules, into whatever ships next.

Constructs As Teammates

A construct is a kind of portfolio piece: portable evidence of how someone thinks about a job. A recurring agent is what happens when the portfolio piece starts showing up to work. A Senior Code Reviewer construct dropped into a NanoClaw nightly job is a named teammate with continuity, a Slack handle, a memory of what it said yesterday, and the same voice tomorrow unless someone forks the construct and changes its mind.

The harness is the engine. The construct is the score. Most of what makes the agent useful on the hundredth run is not in the engine.

What This Does Not Solve

The pattern above is a construct from the registry dropped into a third-party harness that already does the scheduling. NanoClaw is the strongest current example. The same shape works for any custom Claude Agents SDK runner and any self-hosted setup that reads a CLAUDE.md.

Second caveat: writing a recurring-agent persona is not the same skill as writing a one-shot prompt. It has to survive Mondays where there is nothing to say. It has to refuse to fill space. That is a craft question, and the registry is where the craft accumulates.

The Bet on the Other Side of the $12M

Valley Capital Partners did not buy a feature. They bought a thesis about how AI assistants stop being chatbots and start being coworkers. Coworkers have schedules. Coworkers have names. Coworkers have opinions you grow used to and lean on and occasionally argue with. The harness is what gives them the calendar. The construct is what gives them the voice.

NanoClaw is the press. The persona is the writing. The agents are about to start shipping themselves, weekly, on a schedule, and the construct is the only part of them that stays yours.

Related: The Cambrian Explosion of Instruction Files, Teaching Machines to Have Opinions, Constructs Are the New Resume.