2021: GitHub Copilot. Single-line autocomplete. Impressive on standard patterns. Unreliable on custom logic. The model had no awareness of the codebase beyond the current file.
2023: Chat-based coding. Describe what you want. Receive a complete function. Still bounded by what you explicitly provided.
2024: Multi-file awareness. Cursor, Aider, Copilot Workspace. Coordinated changes across files. Refactoring at scale becomes feasible.
2025: Full codebase agents. Claude Code, Gemini CLI, Devin. Describe the goal. Set the agent to work. Review what it produces. Claude Code: 87.6% on SWE-bench Verified, 9 out of 10 real GitHub issues resolved autonomously.
2026: Agent Teams. Multiple Claude sessions operating as a coordinated team. One lead. Specialised teammates. Lateral communication between agents in real time. Concurrent work across the codebase with built-in conflict resolution.
And then there's OpenClaw.
Born in November 2025 as "Clawdbot" by Austrian developer Peter Steinberger. Renamed to "Moltbot" after Anthropic raised trademark concerns. Renamed again to "OpenClaw" three days later because, and I love this: "Moltbot never quite rolled off the tongue."
250,000 GitHub stars. 47,700 forks. In four months.
What made it viral?
It doesn't live in your IDE. It runs on your machine, continuously, and receives instructions through WhatsApp, Telegram, or Discord.
You send it a message at 11pm: "monitor the repo overnight, fix any failing CI tests, open a PR."
You wake up to a PR.
That is a different relationship with an AI agent than anything an IDE plugin can produce. Not because it's more capable per se — but because it's always available, and you never had to open a terminal to talk to it.
The engineering use case: 565+ community skills via ClawdHub. A Cursor Agent skill that delegates implementation tasks to Cursor while OpenClaw coordinates at a higher level. Task Brain: "Kubernetes container scheduling applied to AI agent tasks."
The security reality (see Day 7): it is extraordinarily powerful. It requires the governance posture to match.
The primary failure mode you must know:
The confident wrong answer. Code that is statistically consistent with correct code, but wrong in the specific context of your system. It compiles. It may pass tests. It will surface in production.
It concentrates in: security boundaries, external API integrations, domain-specific business logic, and any code that relies on institutional knowledge not in your specification.
Target your review at these. Trust the standard patterns. Scrutinise the edges.
Tomorrow: the upstream phase that most teams are getting completely wrong, and where AI provides the most leverage that nobody is capturing.
