Agents become teams: myco, friday and Mikka wire coding agents together while lot-map and code-review-gate watch them work

The repo describes itself as “Connect Claude Code, Codex and any agent CLI across machines. Agents find each other, delegate tasks and reply through one hub. Single Go binary, one-line setup.” It is MIT-licensed, created Nov 5, and had 75 stars at the time of our search.

Why it matters now: most people already run more than one agent CLI, and the missing piece is a way for them to hand work to each other without a human copy-pasting. A hub model makes the delegation path explicit and inspectable. The open question is trust: what a delegated agent may do on another machine. We quote the description and did not run it.

The description: “Your coding agents, working together as one team. Claude Code, Codex, Cursor and 13 more talk to each other, split the work on a shared board and check each other's work, each in its own git worktree.” MIT, 62 stars, runs on macOS, Windows and Linux per the repo.

Why it matters now: the git worktree per agent is the quiet design choice. It gives every agent an isolated checkout, so parallel work stays mergeable and reviewable. “Check each other's work” is the claim to test, because agents reviewing agents only helps if the reviewer is independent of the author. Star count is small, so treat this as an idea to watch.

Per the repo: “Move your coding agents off your laptop. Mikka runs each task on its own isolated cloud machine, connects it with the agents on your computers, and sets everything up from one sentence to the Coordinator.” It is AGPL-3.0 and had 84 stars when we searched; it was pushed on Nov 12.

Why it matters now: long-running agents tie up a laptop and run with your credentials. One machine per task trades convenience for containment, the same bargain this week's other sandbox projects make. The “Coordinator” is the interesting layer, since that is where task routing lives. We did not test the setup claim.

The description: “Your own team of AI bots in one self-hosted binary: spaces, every account you have, handoffs, approvals and routines. Open-source, self-hosted alternative to Grok Bot, Claude Duet and OpenAI dots. Rust + React.” It had 506 stars, AGPL-3.0, and was pushed Nov 12.

Why it matters now: it names approvals and handoffs as first-class features next to routines, which is what teams ask for once bots start acting on shared accounts. The product names it positions against are the repo's own wording and we have not verified them. One binary and self-hosting make it easy to try and easy to audit.

The repo pitch: “A live map beside the chat that shows what Claude Code is doing: tool calls, model steps, hooks, skills and subagents light up as they happen. Local only, in five languages.” MIT, 394 stars, pushed Nov 12.

Why it matters now: with hooks, skills and subagents all firing, the chat transcript no longer shows what actually happened. A live view of tool calls and subagents is a cheap answer to “why did it do that?” Running local-only matters because the data is your sessions. We did not run it.

The description: “Auditable code review skill for Claude Code and Codex: four independent probes, evidence-backed findings, and a JSON-derived merge gate.” It is MIT with 472 stars.

Why it matters now: agent-written diffs are arriving faster than people can read them, so review has to become a gate with evidence rather than a vibe. Four independent probes and a JSON-derived verdict give a machine-checkable pass or fail. Whether the probes are truly independent is the claim a reader should check in the source.

Per the repo: “Fault-injected tool environments with side-effect grading for evaluating and post-training LLM agents.” MIT, 104 stars, last pushed Nov 9.

Why it matters now: agents that pass a happy-path demo often fail when a tool times out or returns garbage, and the damage is in the side effects, not the final answer. Grading side effects is the right measure for agents that touch real systems. This is a research-flavoured repo with a small following, so we flag it as a method to watch.

The description: “Write an app once. It runs natively on the web, macOS, iOS, and Linux, and an AI agent can build it, run it, see it, and test it on every one of them.” It is MIT-licensed, written in Rust, and had about 2,900 stars at the time of our search. It lives under the expo organisation.

Why it matters now: “build it, run it, see it, test it” is the full agent loop, and a cross-platform toolchain that exposes all four steps removes the manual screenshot-and-paste step. Attention is high for a repo this new, so expect fast change. We have not built anything with it.

The description: “Your agent finally sees the Compose UI it builds. Jetpack Compose and Compose Multiplatform Previews in your browser, redrawn a second after every save — no Android Studio, no build.” Apache-2.0, 99 stars, from the kakao organisation.

Why it matters now: agents write UI blind unless something renders it for them. A browser preview that redraws on save gives an agent a fast visual check without an IDE or a full build, which is the cheap version of the loop above. It is Android-specific and early; we did not run it.

Per the repo: “Self-hosted Chrome sessions for you and your AI agents. MCP, live human takeover, saved profiles, drag-and-drop files, downloads and audio.” AGPL-3.0, 61 stars.

Why it matters now: browser agents stall on logins and captchas, and the practical fix is a persistent profile plus a way for a person to step in mid-session. “Live human takeover” over MCP is that fix, self-hosted so cookies stay with you. Small star count; treat it as a pattern, not a product verdict.

The description: “The workflows and skills the MiniMax Code team uses to build MiniMax Code. Open source, install and use.” MIT, 698 stars, from the MiniMax-AI organisation.

Why it matters now: a team publishing the exact skills it uses to build its own coding product is a rare look at working practice rather than a demo. If you build agent workflows, this is reading material as much as software. We did not inspect the skills individually.

Per the repo: “44 agent skills from shipping three live products: decide what to build, build it with Claude Code or Codex, have a second model review it, release it.” MIT, 528 stars.

Why it matters now: it covers the whole path from deciding what to build to release, and it bakes in a second-model review step, the same independent-reviewer idea showing up elsewhere in this brief. The 44-skill count is the repo's own claim.

The description: “An agent skill that draws animated icons in one consistent style, for Claude Code, Codex and other coding agents. Comes with 2,600+ open-source icons for React, Vue and SVG, and a website to browse them.” It had 736 stars; the repo's license field reads NOASSERTION, so check the terms before reuse.

Why it matters now: design consistency is where agent output usually drifts, and a skill that pins one style plus a large icon set is a practical constraint. The licence ambiguity is worth a look before shipping icons from it.