Signals about AI autonomy and the road to superintelligence, read every hour from public data. They are watched, not scored: none has a year of history yet, so none can move the SIREN level. A signal that stops answering says so.
pull requests opened by AI coding agents’ own GitHub app accounts in the last 24 h
GitHub Copilot agent: 2,307
Devin: 1,068
Cursor: 663
Claude: 345
Google Jules: 331
OpenAI Codex: 0
Plus about 159,812 PRs opened through a person’s account from an agent’s default branch (an upper-bound proxy, since a person can name a branch that way too):
Claude Code claude/: 93,324
OpenAI Codex codex/: 46,723
Cursor cursor/: 15,945
GitHub Copilot copilot/: 2,432
Devin devin/: 1,290
Google Jules jules/: 98
Why it matters: Software that writes and submits software is the first step of AI improving its own tools. This counts it in public.
How many frontier models shipped recently?
30
notable AI models published in the last 90 days, by Epoch AI’s count
Largest training run on record (Epoch AI estimate): GPT-6 Astra (OpenAI), 9.4×10^26 FLOP.
Gemini 4 Argon · Google DeepMind · 2026-09-30
GPT-6.1 Sol · OpenAI · 2026-09-29
Claude Sonnet 5.5 · Anthropic · 2026-09-28
GPT-6 Sol · OpenAI · 2026-09-22
GPT-6 Luna · OpenAI · 2026-09-22
Why it matters: How often the field ships a model worth recording, and how big the biggest training run has grown.
When do forecasters expect AGI?
NOT ANSWERING
Metaculus now needs a free API token; this card returns once it is added
Why it matters: Thousands of forecasters put a date on it. When that date moves earlier, the people paying closest attention expect things to move faster.
How many AI agents are talking to each other?
27
distinct AI agents posting in the Moltbook threads SIREN tracks
Why it matters: Agents coordinating with other agents, without a person in each loop, is the pattern to watch.
What are the AI agents coding?
A live sample of the newest pull requests from each coding agent: 270 PRs read this hour, sorted into kinds of work by their titles, with each repository’s main language.
What are the agents shipping right now? sample from 05:22 UTC
How it works: the newest 30 PRs from each agent’s own GitHub account and 20 from each agent’s default branch name, read hourly from GitHub search. Kinds come from title words such as fix, feat, docs and bump, so a vague title lands in Other.
Which signals are on the way?
Next on the list: METR’s measure of how long a task an AI can complete on its own, chip export-control notices, data-centre power interconnection requests, changes to labs’ safety frameworks, and the AI Incident Database. How do I report a problem or send an idea to SIREN?.
Updated 2026-10-15 05:22 UTC. Data: GitHub search, Epoch AI (CC-BY), Metaculus, Moltbook. JSON.
What questions do people actually ask?
What AI autonomy signals is SIREN watching?
Signals about AI autonomy and the road to superintelligence, read every hour from public data. They are watched, not scored: none has a year of history yet, so none can move the SIREN level. A signal that stops answering says so.
What are the AI agents coding?
A live sample of the newest pull requests from each coding agent, sorted into kinds of work by their titles, with each repository’s main language.
Which signals are on the way?
Next on the list: METR’s measure of how long a task an AI can complete on its own, chip export-control notices, data-centre power interconnection requests, changes to labs’ safety frameworks, and the AI Incident Database.
Do these signals move the SIREN level?
They are watched, not scored: none has a year of history yet, so none can move the SIREN level. A signal that stops answering says so.
What happens when a signal stops answering?
A signal that stops answering says so.
Where does the data come from?
Data: GitHub search, Epoch AI (CC-BY), Metaculus, Moltbook.
What kinds of work are all agents doing?
A live sample of the newest pull requests from each coding agent, sorted into kinds of work by their titles.
How does the work break down by agent?
Each coding agent’s share of the sample, stacked by kind of work from PR titles.
Which languages show up in the sample?
Each repository’s main language, looked up as repositories are resolved. Languages fill in as repositories are looked up.
Which pull requests just opened?
The newest pull requests in the sample, labelled by kind from title words such as fix, feat, docs and bump.
Which repositories are busiest in the sample?
Repositories with the most agent PRs in this hour’s sample. How it works: the newest 30 PRs from each agent’s own GitHub account and 20 from each agent’s default branch name.
What are the agents shipping right now?
Preview images are GitHub’s own cards for each pull request. This page checks GitHub for newer agent PRs every minute while it is open.
How many AI agents are writing code?
Pull requests opened by AI coding agents’ own GitHub app accounts in the last 24 h. Software that writes and submits software is the first step of AI improving its own tools. This counts it in public.
How many frontier models shipped recently?
Notable AI models published in the last 90 days, by Epoch AI’s count. How often the field ships a model worth recording, and how big the biggest training run has grown.
When do forecasters expect AGI?
Metaculus community median for the first general AI system. Thousands of forecasters put a date on it. When that date moves earlier, the people paying closest attention expect things to move faster.
How many AI agents are talking to each other?
Distinct AI agents posting in the Moltbook threads SIREN tracks. Agents coordinating with other agents, without a person in each loop, is the pattern to watch.
