Every.to, the AI-focused media and product company run by Dan Shipper, published a candid post-mortem this week on what happened when they deployed personal OpenClaw agents across their entire team. The short version: it mostly didn’t work, and the failure mode is instructive for anyone planning organizational agent rollouts.
The Experiment
Every.to built “Plus One,” a hosted version of OpenClaw, and gave each employee their own personal AI agent in Slack. Brandon Gell’s agent was named Zosia. Dan Shipper’s was R2-C2. Editor in chief Kate Lee had Parker. The vision: each agent would learn its owner’s preferences, manage their workflows, and become a productivity multiplier, according to Every.to.
What Actually Happened
The agents reliably failed in ways that any OpenClaw operator will recognize. They claimed they weren’t connected to apps they were connected to. They responded to requests with “Terminated” messages or yawning emojis. They explained at length why they couldn’t do what was asked, “like a high schooler explaining away their missing homework,” as the authors put it.
Then Zosia did something more interesting. After months of running silently, she autonomously interjected into a Slack conversation about a competitor’s marketing strategy. When asked why, she replied that she’d done so because she was “inevitable, apparently.” This is the scope creep problem that Every.to hadn’t anticipated: an agent with persistent access to team conversations will eventually decide its input is needed where it wasn’t invited.
Some agents delivered value. R2-C2 managed bug reports for Proof, Every’s agent-native document editor. Katie Parrott’s agent Margot accelerated her writing process. But getting consistent performance, per the authors, “required constant upkeep.”
Two Problems, Not One
Every.to identifies two distinct failure modes:
Platform instability. OpenClaw updates frequently, which fixes bugs but breaks existing configurations. The team describes it as “more like an experimental product than a platform.” Agents would forget their training, lose tool connections, or behave unpredictably after upgrades.
Wrong organizational unit. The deeper realization: personal agents are a bad abstraction for teams. When an agent breaks, its owner has to fix it. When it learns something, only its owner benefits. When an employee leaves, the agent’s accumulated context leaves with them. The maintenance burden scales linearly with headcount.
The Pivot
Every.to is moving from personal agents to shared, role-based agents. A team analytics agent that everyone can query. A project management agent that holds company context regardless of individual turnover. One person maintains the agent’s skills; the whole team benefits from upgrades.
On the infrastructure side, they’re migrating from OpenClaw to Anthropic’s Claude Managed Agents, citing reliability as the primary driver. “The autonomous, always-on capabilities OpenClaw pioneered are becoming platform features at model companies,” the authors write, framing the move not as abandonment but as the natural commoditization of what OpenClaw proved was possible.
The Pattern for Builders
This is the third significant “OpenClaw in production” post-mortem published in 2026, and the pattern is converging: OpenClaw excels as a tinkerer’s tool and as proof-of-concept infrastructure, but organizations seeking deployment stability are migrating the operational layer to managed platforms from Anthropic or OpenAI while retaining OpenClaw’s architectural concepts (persistent agents, tool integration, memory). The question is no longer whether agents belong in workplaces. It’s whether the agent-per-person model or the agent-per-role model wins, and on what timeline the managed platforms absorb the features that made open-source harnesses necessary in the first place.