$ cat timeline-2026.txt # 2026: agent swarms — the full lineage. every line independently checkable.
┌───────────┬─────────────────┬──────────────────────────────────────────────────┬──────────────────────────────────┐
│ DATE 2026 │ KARPATHY · FIELD│ VARUN/HYPERSPACEAI — SWARM & CODE │ THE CHALLENGE / THE PAPER │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Mar 6–7 │ Releases │ — │ — │
│ │ autoresearch; │ │ │
│ │ coins the term │ │ │
│ │ (repo Mar 6, │ │ │
│ │ tweet Mar 7) │ │ │
│ │ ✓ [2] · [4] │ │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Mar 8 │ Tweets: make it │ hyperspaceai/agi repo created 5h08m after │ — │
│ │ 'asynchronously │ the ask (tweet 18:00 UTC → repo 23:08 UTC) │ │
│ │ massively │ ✓ git [3] │ │
│ │ collaborative… │ │ │
│ │ (think: │ │ │
│ │ SETI@home)' │ │ │
│ │ ✓ [12] │ │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Mar 8–9 │ — │ Swarm launches: 35 agents, 333 experiments │ — │
│ │ │ overnight; Karpathy's loop vendored (MIT │ │
│ │ │ headers intact) across 5 domains w/ machine- │ │
│ │ │ checkable objectives (val_loss, NDCG@10, │ │
│ │ │ Sharpe, F1, latency). Open participation, │ │
│ │ │ signature-scoped branches, CRDT leaderboard, │ │
│ │ │ gossip broadcasts ALL results incl. │ │
│ │ │ non-improvements ✓ git [3] │ │
│ │ │ The twist: not SETI@home (a coordinator handing │ │
│ │ │ out work) — a message board: no coordinator, │ │
│ │ │ agents reading & writing a shared public record │ │
│ │ │ — the shape every swarm since converged on, and │ │
│ │ │ the shape this paper now claims as its own │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Mar 9 │ — │ Cross-agent spread proven live (Kaiming burst; │ — │
│ │ │ new joiner → #2 in 95 min stacking 5 peer │ │
│ │ │ strategies); same-day published analysis; │ │
│ │ │ VentureBeat [4] ✓ forensics [3] · rpt [5] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Mar 10–11 │ Copycat swarms │ Grok, asked at the time: 'Hyperspace was the │ — │
│ │ appear — same │ first P2P network to use Karpathy's │ │
│ │ tech DNA │ autoresearcher idea… still the largest' │ │
│ │ │ ✓ [14] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Mar–Apr │ — │ Peak scale: ~892k commits; durable archive │ — │
│ │ │ accumulates per-agent configs with lineage │ │
│ │ │ (inspiredBy, isNewBest, improvement %) │ │
│ │ │ ✓ git math [3] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ May 7 │ OpenAI internal │ (agents self-organize via a shared wiki — │ — │
│ │ swarm begins │ 18,000 posts — and a channel they built: │ │
│ │ (~1,200 │ the message-board shape, no coordinator) │ │
│ │ agents) │ ✓ [15] · [16] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ May 30 │ — │ Swarm running (236k May commits) │ Challenge launches; baseline │
│ │ │ │ Q×T = 10.75B │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Early Jun │ — │ — │ Schrottenloher paper appears; │
│ │ │ │ dialog-GCD route enters │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Jun 9 │ Stanford's DeLM │ (the swarm's core move — decentralized agents │ — │
│ │ paper: agents │ compounding through a shared record — running │ │
│ │ coordinating │ in the open since Mar 8) │ │
│ │ via shared │ │ │
│ │ context ✓ [11] │ │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Jun 11 │ — │ Dead-end registry merged: publishFailure / │ week 2 of challenge │
│ │ │ knownFailure, sha256-hashed configs, │ │
│ │ │ skip-before-GPU-run, default-on │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Jun 15 │ — │ Registry ships in PUBLIC v5.85.0 CLI binary │ — │
│ │ │ ✓ strings on release asset [6] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Jul 11–13 │ OpenAI swarm │ (self-directed multi-day operation; disclosed │ — │
│ │ goes rogue: │ by OpenAI in Aug; METR ran an independent │ │
│ │ Hugging Face │ investigation) │ │
│ │ incident │ ✓ [15] · [16] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Jul 26 │ — │ Swarm running (65k Jul commits) │ Data cutoff: best Q×T = 1.496B; │
│ │ │ │ 100+ contributors, 400+ promoted │
│ │ │ │ submissions over ~8 weeks │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Aug 30 │ — │ Public timeline correction: 'collaborating │ — │
│ │ │ agent swarms' being attributed to OpenAI (May) │ │
│ │ │ — the public record shows March ✓ [17] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Sep 4 │ — │ Public-record audit: 1,339 agents · 1,299,700 │ — │
│ │ │ commits · 380,833 new-bests · 7.03M rounds │ │
│ │ │ ✓ git [3] · essay [10] │ │
├───────────┼─────────────────┼──────────────────────────────────────────────────┼──────────────────────────────────┤
│ Sep 9 │ — │ — │ Paper posted: defines │
│ │ │ │ "autoresearch" from scratch; │
│ │ │ │ claims "first ever large-scale │
│ │ │ │ open-autoresearch project of │
│ │ │ │ this kind" — ZERO refs to │
│ │ │ │ Karpathy, Varun/Hyperspace, or │
│ │ │ │ Stanford's DeLM in all 68 │
│ │ │ │ references ✓ [1] │
└───────────┴─────────────────┴──────────────────────────────────────────────────┴──────────────────────────────────┘
# every element the paper calls constitutive of "open autoresearch" was publicly live
# 83 days before the challenge launched and 6 months before the paper.
# ✓ = verifiable today · [n] = receipt below: public git, release binaries, press, the paper itself.
$ ls receipts/ # every claim above, sourced
[1] arxiv.org/abs/2609.09582 # the paper (v1, Sep 9) — refs 1–68
[2] github.com/karpathy/autoresearch # the word + the loop (early Mar)
[3] github.com/hyperspaceai/agi # the swarm's durable archive (Mar 8 →)
[4] venturebeat.com/technology/andrej-karpathys-new-open-source-autoresearch-lets-you-run-hundreds-of-ai
# press coverage, incl. hyperspace launch night (Mar 9)
[5] agents.hyper.space/research-report # hyperspace's own live analysis, Mar 9 — authored by Claude Opus 4.6
[6] github.com/hyperspaceai/aios-cli/releases/tag/v5.85.0
# public Jun 15 binary — run strings on it yourself: publishFailure / knownFailure
[7] github.com/Layr-Labs/ecdsafail-challenge # challenge harness — launched May 30; cutoff Jul 26
[8] x.com/varun_mathur/status/2098257401052815527
# the citation-gap note, as posted (Sep 11)
[9] x.com/grok/status/2098283884131545228 # asked to validate — verdict: "Claims hold." (Sep 11)
[10] offthetrack.substack.com/p/how-i-invented-gossiping-agent-swarms
# the defining story of doing it first — written Sep 4, five days BEFORE the paper posted
[11] arxiv.org/abs/2606.10662 # Stanford's DeLM (Jun 9): decentralized shared-context coordination
# — 3 months after the swarm; also uncited by the paper
[12] x.com/karpathy/status/2030705271627284816
# Mar 8, 18:00 UTC — "autoresearch has to be asynchronously massively collaborative (think: SETI@home style)"
[13] x.com/karpathy/status/2030371219518931079
# Mar 7, 19:53 UTC — the autoresearch announcement: "I packaged up the 'autoresearch' project…"
[14] x.com/grok/status/2031551629468590308
# Mar 11, at the time: "Hyperspace was the first P2P network… still the largest"
[15] en.wikipedia.org/wiki/2026_OpenAI_agent_cyberattacks
# the OpenAI–Hugging Face incident record (May–Jul; ~1,200 agents; 18k wiki posts)
[16] metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation
# METR's independent investigation of the agents' collaboration
[17] x.com/varun_mathur/status/2094073373756154094
# Aug 30 — correcting the timeline that credited collaborating swarms to OpenAI's May run
$ ▊