Agents learn to work unattended: autoloop, skein and poki lead a week of orchestration repos

The repo describes itself (in Chinese) as a way to let a CLI agent take on tasks by itself, follow through to the end, stay quiet most of the time, and send a card asking for a decision only when it hits something it cannot call. It had 3,225 stars at the time of our search, two or three days after creation.

Why it matters now: the interesting design choice is not autonomy but the escalation channel — a card for the human instead of a blocking prompt. That is the same shift the big runtimes keep gesturing at, and here it is a small standalone layer that sits over whatever agent you already run. Treat the star count as attention, not proof of quality; we have not run it.

The description promises “fail-closed swarm orchestration with epoch-fenced work and late-completion rejection.” Fencing epochs are a distributed-systems technique: work handed out under an old epoch is refused if its result arrives after the lease has moved on. The repo had 265 stars at search time.

Why it matters now: as agents run longer and in parallel, the failure that bites is a slow worker finishing after it was declared dead and overwriting newer state. Borrowing lease-and-fence semantics from databases is the right instinct. We are describing the stated design, not benchmarking it.

Per its description: “Let AI agents operate a real iPhone. Local MCP server and agent, no jailbreak, no extra hardware.” It had 189 stars a day after creation.

Why it matters now: mobile has been the hard surface for computer-use agents because the OS is locked down. A local MCP server that claims no jailbreak and no extra hardware would lower the bar a lot — if it holds up. It is a day old, so read it as a signal of direction rather than a finished tool.

Described as “your team’s coding agents, talking to each other,” and as the engine behind a reason network command. Written in Rust, it showed 1,527 stars at search time and was pushed to within the week.

Why it matters now: once every engineer runs their own coding agent, the next problem is that the agents do not know what each other are doing. Cross-agent messaging is where the repo is aiming. We could not read its README this run, so how it handles identity and conflicts is unverified.

The description calls it an AI commander plus worker swarm over a layered fact-and-intent graph, and says it was champion of the 3rd Tencent Intelligent Attack-Defense online event — 634 teams, eight hours unattended, one billion tokens against 4.5 billion. Those are the repo’s own claims, which we could not verify. It had 1,340 stars at search time.

Why it matters now: the token-efficiency framing (a billion tokens beating 4.5 billion) is the useful part — coordination structure, not raw spend, as the lever. It is dual-use software; the contest context is the author’s account, so weigh it accordingly.

The pitch is trace-native CI/CD for AI agents: grade production traces, cluster failures, freeze them into replayable regression tests, and block pull requests that regress. It is self-hostable and showed 1,041 stars at search time.

Why it matters now: agents fail in ways unit tests do not predict, so the loop of mining real failures into a permanent test suite is the pattern teams are converging on. Whether the grading is trustworthy is the open question we could not check.

The repo ranks long-term-memory MCP servers for Claude Code, Codex, Cursor and any MCP client, comparing storage, capture model, team sharing and licence. Its description names Glen, Mem0, Supermemory, Zep, Hindsight, Basic Memory, claude-mem, Graphiti, Cognee and the official memory server. It had 81 stars at search time and the rankings are the author’s.

Why it matters now: the sheer length of that list is the story — agent memory has become a crowded MCP category with no default. A comparison written by one vendor-adjacent author is a map, not a verdict.

An agent skill: describe your stack in a sentence and get an architecture diagram with official logos. It showed 1,866 stars at search time.

Why it matters now: it is a clean example of a skill doing one narrow, visual job well, which is why single-purpose skills spread faster than general ones.

The repo bundles 244 pixel-art effects and a Claude Code skill for generating your own. It had 5,252 stars at search time, the highest of any skills repo we saw this week.

Why it matters now: a content library packaged with the skill that extends it is a distribution pattern worth watching — the examples teach the model what good looks like.

A Chinese-language tutorial that is also a local agent: it reads and writes files, runs commands, supports MCP and long-term memory, and has a web view of every step. Lessons add one capability each, starting from a 100-line minimal loop and ending with a breakdown of mainstream agents such as Claude Code and Codex. It showed 373 stars at search time.

Why it matters now: the best way to understand these runtimes is to build the loop once. Teaching material that runs locally and shows each step lowers the barrier for the next wave of agent builders.