Skip to content

Instantly share code, notes, and snippets.

@dims

dims/2026-06-03-substrate-cross-vendor-contributors.md

Created June 4, 2026 13:35
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save dims/dc0c769bd5b94b4ebabecc60dbcd3dc8 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save dims/dc0c769bd5b94b4ebabecc60dbcd3dc8 to your computer and use it in GitHub Desktop.
Download ZIP
Agent Substrate (agent-substrate/substrate) — Cross-Vendor Contributor & Affiliation Report (2026-06-03)
Raw
2026-06-03-substrate-cross-vendor-contributors.md

Agent Substrate — Cross-Vendor Contributor & Affiliation Report

Generated: 2026-06-03 Repo: agent-substrate/substrate — "Agent Substrate: the core system" (public, 468★, Apache-2.0) What it is: a system on top of Kubernetes that manages agent-like workloads at higher scale/lower latency by taking the K8s control-plane out of the critical path — actors run in gVisor sandboxes (ateom), managed by a kubelet-like agent (atelet), with GCS checkpoint/restore (ategcs) and a router (atenet). Window: 2026-05-13 → 2026-06-03 (~3 weeks — a brand-new seed project) Volume analyzed: 95 commits · 117 PRs (all states) · 63 issues (all states) Analysis basis: upstream agent-substrate/substrate@main (e26cfa22), cloned fresh — the local dims/substrate fork checkout (4cbac18) was a few commits behind.

Framing: Unlike the NVIDIA-owned reports (nvsentinel / dra-driver / aicr / OpenShell), this repo is not NVIDIA-owned — it's a Google-seeded, multi-vendor project under the neutral agent-substrate org. So this is an affiliation breakdown across all contributors (commits + PRs + issues), with NVIDIA's footprint called out, rather than an "external = non-NVIDIA" split.

Methodology: git log (authors + DCO trailers) on the upstream clone → gh pr list/gh issue list --state all for PR/issue authorship → resolved logins via gh api repos/.../commits/{sha} → classified each contributor's employer via GitHub company field, commit-email domain, NVIDIA-org membership (gh api /orgs/NVIDIA/members, privileged token), Helios (NVIDIA folks), and targeted web search → merged email aliases.

Governance note (key): substrate uses the Google CLA (cla.developers.google.com), not DCO sign-off. CI is govulncheck + pr-workflow — there is no DCO check. So the very low Signed-off-by rate (8/95 commits) is expected, not a hygiene problem; CLA is the legal gate. Sign-off here is just personal habit (see DCO section).

Affiliation Breakdown

Affiliation Contributors Commits PRs (merged / open / closed) Issues (open / closed)
Google (seed team — owns the core) 12 63 40 / 11 / 3 35 / 9
NVIDIA 2 15 14 / 6 / 4 0 / 1
Solo.io 6 3 3 / 5 / 2 5 / 2
Independent (Arnaud Meukam) 1 4 4 / 1 / 0 0 / 0
Hugging Face (Adrien Delorme) 1 3 3 / 1 / 1 0 / 0
CoreWeave 1 2 2 / 0 / 0 1 / 0
AWS 1 1 1 / 2 / 0 1 / 0
Red Hat (jbpratt) 1 1 1 / 0 / 0 0 / 0
Alibaba Cloud 2 0 1 / 0 / 0 1 / 0
Undisclosed / individuals 7 2 2 / 3 / 3 6 / 0
Bot / App (a4-a4s1) 1 0 0 / 0 / 4 0 / 1
Seed-team group commit — 1 — —
Total ~34 humans 95 67 / 28 / 22 50 / 13

Headline: Google seeds and owns the core (63 of 95 commits, ~66%); NVIDIA is the #2 contributor by commits (15) — almost entirely Davanum Srinivas. The most engaged outside company is Solo.io (6 people, networking/security/architecture), which shows up far larger in issues/PRs than in commits. Alibaba Cloud appears only in issues/PRs (pluggable-backend proposals), not commits — exactly the engagement the "commits-only" view would have missed.

Google — the seed team (12 contributors, 63 commits)

Google authored the repo and owns the runtime/architecture. Notably, the senior leads push directly to main (Hockin: 25 commits but only 7 merged PRs → ~18 direct pushes; Elder: 16 commits / 13 PRs), while outside contributors go through PRs.

  • Tim Hockin — @thockin — 25 commits, 7 merged PRs, 13 issues (10 open). Kubernetes co-founder / SIG-Network & SIG-Architecture. The project's architectural lead — repo bootstrap, conventions, scaling use-cases, and most of the open design issues.
  • Benjamin Elder — @BenTheElder — 16 commits, 13 merged PRs. K8s Steering Committee, SIG-Testing TL, KIND creator. Owns build/CI/vendoring (ko, govulncheck, go upgrades), ateom socket-path limits, OCI untar correctness.
  • Julian Gutierrez Oschmann — @juli4n — 9 commits, 7 merged PRs, 7 issues. atelet on object storage, WorkerPool /scale subresource, repo reorg tooling, code-layout docs.
  • Taahir Ahmed — @ahmedtd — 5 commits, 5 merged PRs, 7 issues. runsc flags, image-pin enforcement, trace IDs; landed "Adopt the Kubernetes AI contribution policy."
  • Haven Xia — @HavenXia — 4 commits + PRs (@google-gemini). Logging refactor, Cloud Monitoring dashboard, kubectl ate logs --follow streaming.
  • Grant McCloskey — @MushuEE — 1 commit, several PRs/issues. Persistence/scale: Aerospike store + benchmark, O(1) scheduler via Valkey/Redis, actor-list pagination.
  • Dmitry Berkovich — @dberkov — 1 commit + 3 issues. Actor state-machine/lifecycle, ActorID injection, the "logoless" bug.
  • Max Smythe — @maxsmythe — issues only (5 open). Google; Gatekeeper/OPA maintainer. Drives the scale & benchmarking agenda (high-scale use cases, observability-at-scale, automated benchmarking, stress testing).
  • Bowei Du — @bowei — issues only (2). Google networking — "NetworkPolicy by default", "harden interior net filtering."
  • Troy Chiu — @troychiu — 1 commit. (@google; @kubernetes/@ray-project/@flyteorg.)
  • Jaana Dogan — @rakyll — 1 commit ("Use Antigravity in README"). Google, well-known eng.
  • Daniel Schwanden — @daschwanden — 1 PR + 1 issue: IaC for GCP resources. (Google Cloud, per work.)

NVIDIA (2 contributors, 15 commits — the #2 affiliation)

  • Davanum Srinivas — @dims — 13 commits, 13 merged PRs (+6 open, 4 closed) — dsrinivas@nvidia.com (Helios-confirmed, hired 2025-07-14, SW-GPU). Version stamping + --version, ateapi/syncer actor-release-on-pod-delete, serverboot extraction/refactor, ateom-gvisor eth0 move idempotency + rollback, CI workflows. The single most DCO-disciplined contributor in the repo (12/13 signed) — see DCO note.
  • Mark Chmarny — @mchmarny — 2 commits, 1 merged PR, 1 issue — mchmarny@nvidia.com (Helios-confirmed). Router error handling (map gRPC → HTTP status codes); filed the macOS local-dev-friction issue.

NVIDIA's footprint is concentrated in the router, the gVisor networking path, and build/release plumbing — the same operational-hardening shape NVIDIA shows on other repos. dims is a genuine top-4 committer here, not a drive-by.

Solo.io — the most engaged outside company (6 contributors)

Heavily focused on networking, security, and architecture proposals (the Istio/Gloo wheelhouse). Shows up far larger in issues/PRs than commits.

  • John Howard — @howardjohn — 2 commits/merged PRs + 1 open. Istio @ Solo.io. Router target-IP logging; K8s 1.36 pod certificate request handling; proposed pluggable networking providers (#85).
  • Yuval Kohavi — @yuval-k — Solo.io co-founder/CTO. Architecture proposals: fold atelet into ateom (#128), run on vanilla K8s (#127), cache JWT/CA instead of disk reads (#17).
  • Eitan Yarmush — @EItanya — 2 open PRs, 4 open issues. Big networking/security push: replace actor eth0-move with worker-owned veth networking (#110/#122), transparent egress policy + credential injection (#126), ConfigMap/Secret env sources, Helm charts for install (#79).
  • Nina Polshakova — @npolshakova — PR + 2 closed issues. ActorTemplate mutation/regeneration bugs, kind cluster setup fixes.
  • Peter Jausovec — @peterj — 1 commit (signed). Sort actors/workers alphabetically.
  • eliranw — @eliranw — PRs: "atelet runs only on nodes hosting ateoms" (#133/#134). (Likely Solo.io — Istio/Gloo ecosystem; unconfirmed.)

Other vendors

Hugging Face — Adrien Delorme @azr — 3 commits + PRs + 1 issue. Ex-HashiCorp (founding Packer release engineer, ~8 yrs), now Hugging Face; Stockholm. Runtime robustness: ateom socket-path validation, GCS upload-error surfacing, move working dir off tmpfs, memory pull-cache failures.

CoreWeave — Brandon Jacobs @brandonrjacobs — 2 commits + 1 issue. Infra @ CoreWeave — GPU clusters. ategcs CreateBucket-per-Put fix, ateom-gvisor route ordering; filed "high concurrent active actors per worker."

AWS — Kyosuke Konishi @konippi — 1 commit + PRs/issue (@aws, Japan). Security + observability: path-traversal fix in OCI tar extraction (#101), Prometheus dev stack, telemetry service.instance.id.

Red Hat — jbpratt @jbpratt — 1 commit (signed). localca unit tests.

Independent — Arnaud Meukam @ameukam — 4 commits (all signed) + PRs. Kubernetes SIG-K8s-Infra co-chair (co-leads it with dims), release manager. CI/infra hygiene: govulncheck workflow, misspell linter, .gitignore, base-image bump, more golangci-lint plugins.

Alibaba Cloud (issues/PRs only — no commits) — ringtail @ringtail (Alibaba Cloud): "Aliyun OSS adapter" PR (#180). Yuanhong Peng @yummypeng (Alibaba, Hangzhou): proposal to make the ateom backend pluggable beyond gVisor (#121) — i.e. swap in their own sandbox (Kata/runD). Signals Alibaba wants substrate to run on their sandbox + object storage.

Undisclosed / individual contributors

  • Michael Taufen — @mtaufen — 3 open issues (security-leaning: log sanitization, mitigate lateral movement via snapshot suspend/resume, multi-substrate operator). Ex-Google kubelet (dynamic config); current affiliation undisclosed.
  • Alex Bulankou — @AlexBulankou — 1 commit + 2 closed PRs (Redmond, WA). Pod-cert-controller graceful skip, deletion finalizer, DNS readme. Employer undisclosed.
  • Maya Wang — @mayawang — 1 commit + PRs/issue (Seattle). Docs (Agent Executor ecosystem), OpenClaw multiplexing demo (#177). Undisclosed.
  • Falven — @Falven — PRs/issues: configurable router timeouts, avoid re-untar on restore. Undisclosed.
  • ArnaudBger — @ArnaudBger — atelet per-actor cgroup path (#161). Undisclosed.
  • summersummerwang — @summersummerwang — closed PR: CLAUDE.md + codeexec actor-lifecycle benchmark. Undisclosed.
  • Zoe Zhao — @zoez7 — issue: mTLS auth between all ate-system components (#170). Undisclosed.

Bot/App: app/a4-a4s1 — a GitHub App (4 closed PRs, 1 closed issue), automation, not a human.

DCO / sign-off (expected-low — CLA gates, not DCO)

substrate requires the Google CLA, not DCO. Only 8 of 95 commits carry a Signed-off-by trailer, and they come almost entirely from contributors who carry the habit from Kubernetes:

Contributor Affiliation Signed / commits
Davanum Srinivas (dims) NVIDIA 12 / 13
Arnaud Meukam (ameukam) Independent 4 / 4
Peter Jausovec (peterj) Solo.io 1 / 1
jbpratt Red Hat 1 / 1
everyone else (incl. all Google: Hockin 0/25, Elder 0/16) — 0

The standout is that NVIDIA's dims is the most DCO-compliant contributor on a repo where the Google seed team signs ~0% — a habit difference, not a compliance issue, since the project gates on CLA.

Notable Observations

  1. Google-seeded, but genuinely multi-vendor already at 3 weeks old. 9 distinct companies + several individuals are engaged: Google, NVIDIA, Solo.io, Alibaba Cloud, CoreWeave, AWS, Hugging Face, Red Hat, plus independent K8s-infra leadership. For a repo this young (95 commits), that breadth is unusual and signals a deliberately open, cross-industry seeding.

  2. NVIDIA is the #2 contributor and it's almost entirely one person. dims has 13 commits / 13 merged PRs — real runtime work (gVisor networking, syncer, serverboot, release plumbing), not docs. Mark Chmarny adds the router error-mapping. NVIDIA's stake here is hands-on.

  3. Issues/PRs reveal engagement that commits hide — this is why you asked for them. Solo.io (6 people) and Alibaba Cloud (2) barely register in commits but are among the most active in proposals: Solo.io is pushing the networking/security model (veth, egress policy, credential injection, pluggable network providers, fold-atelet-into-ateom) and Alibaba wants a pluggable sandbox backend + Aliyun OSS. Max Smythe (Google) owns the scale/benchmarking agenda entirely through issues. A commits-only view would have called these "1–2 commit" or "absent" players.

  4. The senior Google leads push straight to main. Hockin landed ~18 commits with no PR; Elder similar. Outside contributors (NVIDIA, Solo.io, Alibaba, etc.) go through PRs. Typical of an early Google-seed repo before formal review gates harden.

  5. A who's-who of Kubernetes is on this repo. Tim Hockin (K8s co-founder), Ben Elder (Steering/KIND), Jaana Dogan (rakyll), Max Smythe (Gatekeeper), John Howard (Istio), Yuval Kohavi (Solo.io CTO), Arnaud Meukam + dims (SIG-K8s-Infra co-chairs), Adrien Delorme (ex-Packer). substrate has attracted unusually senior cloud-native talent fast.

  6. Security is a live, multi-party thread. Path-traversal fix (AWS/konippi), lateral-movement mitigation + log sanitization (mtaufen), mTLS between components (zoez7), credential injection + egress policy (Solo.io/EItanya), image-pin enforcement (Google/ahmedtd). For an agent-sandbox runtime, the sandbox-escape/security surface is getting early attention from several orgs.

  7. CLA, not DCO — and it's a Google CLA. CONTRIBUTING.md points at cla.developers.google.com. This is the clearest signal of who's steering: it's legally a Google-CLA project even under the neutral agent-substrate org name. Worth noting for anyone tracking governance/IP posture.

  8. Naming hints at the lineage. Commits reference OpenClaw (the OpenShell predecessor; Maya Wang's multiplexing demo) and Antigravity (Google's agent IDE; rakyll's README PR). substrate sits in the same "run AI agents as checkpointable actors" space as NVIDIA's OpenShell — which is why it shows up in your dims/* fork sweep alongside openshell.

Methodology Appendix

Commands: git clone https://github.com/agent-substrate/substrate (local fork was behind) → git log --format='%H\t%ae\t%an\t%aI\t%(trailers:key=Signed-off-by,valueonly)' → gh pr list --repo agent-substrate/substrate --state all --json number,state,author,title and gh issue list ... --state all → gh api repos/agent-substrate/substrate/commits/<sha> --jq .author.login → gh api /users/<login> (company/location/bio) → gh api /orgs/NVIDIA/members/<login> (privileged token) → helios-cli user get for NVIDIA folks → WebSearch for Meukam (Independent), Smythe (Google), Delorme (Hugging Face).

Classification: by employer, via GitHub company + commit-email domain + NVIDIA-org/Helios (NVIDIA) + web. Alias merges: juliangut@google.com+julian.gutierrez.o@gmail.com → juli4n; taahm@google.com+ahmed.taahir@gmail.com → ahmedtd; davanum@gmail.com → dims (NVIDIA); mark@chmarny.com → mchmarny (NVIDIA).

Caveats:

  • Counts are authored commits/PRs/issues; PR reviews and issue comments (another engagement axis) were not tallied.
  • Local checkout (dims/substrate@4cbac18) was behind upstream; analysis used upstream main@e26cfa22 (95 commits).
  • eliranw (likely Solo.io), daschwanden (likely Google Cloud), and zoez7 employers are inferred/unconfirmed and flagged as such; summersummerwang, Falven, ArnaudBger, mayawang, AlexBulankou, mtaufen left "undisclosed" rather than guessed.
  • gh api /orgs/NVIDIA/members run with an NVIDIA-employee token (sees private members) — confirmed only dims + mchmarny are NVIDIA.

Run summary (counts)

  • 95 commits · 117 PRs · 63 issues · ~34 human contributors across 9 companies + independents
  • Google 63 commits (66%) · NVIDIA 15 (#2) · all others ≤4 each
  • Most-engaged outside company by issues/PRs: Solo.io (6 people); Alibaba Cloud engages via proposals only (0 commits)
  • DCO: 8/95 signed (CLA-gated, not DCO) — dims most compliant (12/13)
  • NVIDIA confirmed: dims, mchmarny (Helios + org). No other NVIDIA contributors.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment