A recap of what caught our attention at Kaiku this week.
Spec-driven development for agents
OpenSpec landed in the channel this week — a spec-driven development (SDD) framework built specifically for AI coding assistants, aimed at giving an agent a clear, agreed contract to build against before it starts writing code.
Agent sandboxing is getting less painful
Isolating an agent from the rest of your filesystem while still giving it a useful leash has been one of the fiddlier parts of running coding agents, and two tools this week suggested that’s improving. Docker Sandboxes has matured noticeably since a first look back in spring — the isolated microVMs are more solid, and it now ships monitoring, governance profiles, and an MCP gateway of sorts. It’s not polished everywhere yet (auth setup with some coding agents is still a bit of a slog), but it’s a clear step up. agentexecutor.io earned a similarly warm reaction this week for automating sandbox setup entirely, which has otherwise been a genuine pain to configure by hand.
Agents that hire humans
RentAHuman lets AI agents hire people directly through MCP and REST APIs — effectively flipping the usual “human hires an agent” relationship on its head. The discussion in the channel focused on the security angle: what it actually means to give an agent access to money and the ability to transact with a person on the other end, with the obvious follow-on questions about oversight and abuse.
Anthropic moves into project-management territory
Anthropic’s rework of its Projects feature pushes it further into space currently occupied by tools like Linear. Worth watching over the coming months to see where the line settles between “AI assistant workspace” and dedicated project-management tooling.
Want to talk about where AI-native delivery is heading? Get in touch.