AI Agent GitHub Digest — 2026-10-26

🌏 中文版

Today's Highlight

Today's trending repos are all wrestling with the same problem: how do you keep agents from falling apart once there are a lot of them? PrimeIntellect builds "programmable sub-agents plus bounded self-improvement" into a single coding harness. Microsoft Research goes the other way, proving you can scale to over a thousand agents by removing the central coordinator rather than adding one. And an independent developer's WhatsApp MCP shows what it looks like to be honest about the encryption model and the trust boundary when an agent tool reads your private data. In the background, the new Claude Code release makes monitoring itself a first-class feature.

Trending Repos

PrimeIntellect-ai/prime-agent ⭐ 22,325

GitHub　·　Rust　·　MIT

• What it is: An open-source coding and research agent from Prime Intellect, built around the Recursive Language Model (RLM) — context is treated as variables, and tools and sub-agents as function calls, all running inside a persistent Python REPL. A "Continual Harness" lets the agent run /refine to write small, evidence-backed updates from what it learned mid-run — lessons, memories, skill descriptions — back into durable state.
• Why it matters: Most coding agents bolt memory on from the outside, via RAG or scratch files. Prime Agent makes "programmable sub-agent calls" and "scoped self-improvement" first-class instead: it never rewrites the immutable base system prompt, and every update leaves a rollback-able snapshot. A daemon keeps sessions alive across disconnects, and agents can message each other directly without routing everything back through the user — solving the familiar problem of a long-running agent task dying the moment you close the terminal.
• Tech stack: Rust core + persistent Python REPL + RLM (paper: arXiv:2608.23552)
• Getting started: Low — a one-line install script. But it runs model-generated Python with your own user permissions, so the maintainers recommend a disposable clone or restricted environment rather than treating it as a sandbox.

microsoft/Agensh ⭐ 238

GitHub　·　Python　·　MIT

• What it is: Microsoft Research's open-sourced "decentralized, scalable multiagent harness," which scales to as many as 1,024 agents with no central orchestrator. Agents self-organize to claim sub-tasks, share findings, verify each other's results, and merge progress asynchronously, connected through a shared workspace (Gitea), a messaging layer (Mattermost), and a shared-context mechanism adapted from DeLM.
• Why it matters: Most multi-agent frameworks hit a coordination-cost wall after a few dozen agents. Agensh's paper pushes straight to 1,024 and quantifies what coordination itself buys you: on the five hardest ProgramBench tasks, using GPT-5.6-sol with high reasoning effort, a six-hour budget and no internet access, the mean final test-pass rate rises from 19.31% with one agent to 28.78% with 128 agents — roughly a 49% relative improvement — and the Pandoc task reaches 55.06% at 1,024 agents. With the newer GPT-6.1-sol, the six-hour Pandoc pass rate climbs from 56.19% with one agent to 79.59% with 32 agents, about a 41.64% relative gain.
• Tech stack: Python + Gitea (shared workspace) + Mattermost (messaging) + the Copilot harness (the per-agent execution loop integrated for these experiments)
• Getting started: High — this release is a single-node reproduction setup (a Docker Compose multi-agent team plus a standalone single-agent baseline, both on native Copilot ACP sessions), meant to reproduce the paper's experiments rather than to run in production.

finereli/whatsapp-mcp ⭐ 375

GitHub　·　TypeScript　·　MIT

• What it is: A read-only MCP server that lets Claude, ChatGPT, or Grok search and read a user's own WhatsApp messages. You can use the maintainer's free hosted beta, or deploy your own copy on your own Cloudflare account.
• Why it matters: Most WhatsApp integrations either hand your account to a third-party cloud service or say nothing about encryption. This project is unusually explicit about all three pieces: the connection model (it behaves like another linked WhatsApp Web device, connecting only when the agent asks and disconnecting right after), the encryption model (one Cloudflare Durable Object per person, a data key sealed with AES-256-GCM, phone numbers and chat IDs replaced by keyed hashes), and exactly who can see what — including the honest admission that the hosted version's login-service operator can, in principle, derive your key, and that this is not end-to-end encryption.
• Tech stack: Cloudflare Workers + Durable Objects (with embedded SQLite) + Baileys (a WhatsApp Web protocol library) + the MCP TypeScript SDK
• Getting started: Medium — the hosted version just needs a QR-code scan, but self-hosting requires a Cloudflare Workers Paid plan (from $5/month) and the project has to patch around three compatibility gaps Baileys has on the workerd runtime (WebSocket, wasm loading, timers).

Notable Releases

Claude Code v2.1.310

Release Notes

• Key changes: Added a Monitor tool and plugin monitoring, covering Bedrock, Vertex, and Foundry as well as sessions with telemetry or non-essential traffic disabled, and extended the same restriction to long foreground sleeps. Fixed requests failing behind some gateways and compatible servers that reject the effort-2025-11-24 beta value or mid-conversation system messages — Claude Code now retries without them. Fixed compaction after a "Prompt is too long" error keeping too much dense text (such as base64 or SVG) in context, which could make the next request fail again.
• Breaking Changes: None.
• What it means for you: If your workflow relies on long-running background monitoring, or runs on Bedrock, Vertex, or Foundry, you now have a native Monitor tool. If you've been blocked by certain compatible gateways before, this release should fix that.

Today's Takeaway

I used to think the bottleneck in multi-agent coordination was mostly about model reasoning — if the agents aren't smart enough, coordination won't work well either. Agensh's numbers say otherwise: with the model held completely constant, just swapping the coordination mechanism from central dispatch to self-organized claiming plus asynchronous merging buys a 40-50% lift in pass rate on the same task. The coordination architecture itself may be squeezing more value out of adding agents than waiting for the next model upgrade ever could.

References

• PrimeIntellect-ai/prime-agent — GitHub
• Prime Agent: A Self-Improving RLM Harness — arXiv:2608.23552
• microsoft/Agensh — GitHub
• Agensh: Scaling Organizational Intelligence to 1,024 Agents — arXiv:2609.26781
• finereli/whatsapp-mcp — GitHub
• Show HN: WhatsApp MCP running free on Cloudflare — Hacker News
• Claude Code v2.1.310 Release Notes