Agent-harness race: OpenClaw 2.0 and xAI Grok Build

Closed · 6 updates · since Sep 7, 2026 · updated Oct 29, 2026 AI openclaw xai Elon Musk agents

Where it stands: OpenClaw and xAI's Grok Bot remain in a fast release race: OpenClaw pushed small patch releases, while Grok Bot announced phone-calling, phone-app integrations, and more commerce tools. The unresolved test is whether bots complete real purchases and tasks; Amazon has blocked an unauthorized agent, and Musk floated a rival store. Watch whether Grok Bot calls ship and whether agent-commerce access gets resolved.

Latest

2026-10-03 the funding layer prints a revenue number and a silicon layer appears. (1) Vercel at $600M annualized revenue, +148% YoY; agents ≈ half of new business, up from <3% at start of year (The Information via nichochar RT, ~8:01 PM ET), caption teasing a "fundraise announcement incoming" (X-grade, ledgered) — the agent-led revenue conversion at the dev-infra layer, and Vercel's gateway is already Jev/distribution (decision-models adjunct). (2) Perplexity verticalizes on NVIDIA Vera (Arav, ~5:19 AM ET): "own our sandboxes, and optimize the sandbox for the best silicon. Vera is far better than x86" — Perplexity Computer to deploy on Vera ("SPACE" companion note); bare vendor claim, no numbers/date, rollout "soon" (ledgered). First announced silicon vertical for an agent sandbox. (3) OpenClaw v2026.9.8 (~12:59 AM ET): GPT-6.1 Sol support, agent-reply improvements, less memory, Windows fixes; 43 PRs/8 contributors. Friction (rescued from filtered): steipete — OpenClaw's Android app "over a week in review limbo" on Google Play (~2:07 PM ET) — app-store review as the harness ecosystem's distribution gate. (4) Grok Bot cadence: Lingxi Li solicits two-week-horizon feedback (Elon RT, ~3:13 AM ET) — iteration mode post-verdict.
2026-09-18/19 a wave — (1) OpenClaw 2026.9.5 (fifth release in 19 days: 2.0→9.2→9.3→9.4→9.5; 4,179 PRs/502 contributors; atomic updates, plugin hot reload, conversation sharing, specialist-agent setup) + the Multiplayer Mode deep-dive ("the death of the meat proxy" — agents relaying directly, humans out of the loop) + an OpenClaw × aiworthusing hackathon. (2) Claude Code v2.1.277 adopts AGENTS.md (checks it when no CLAUDE.md exists; toggleable; Dohmke: "AI is finally aligned 😜") — the standing watch item (converge or fork) answers: converge. Mechanics (changelog + The Agents Desk, Sep 19): substitution, not merge — per folder, AGENTS.md is read only when no CLAUDE.md exists (CLAUDE.md wins and suppresses it; no co-reading); on by default, toggleable in /config under "Project instructions"; NOT yet live on Bedrock/Vertex/ Foundry. Closes issue #6235 (6,653 reactions, 13 months, closed as "completed" Aug 17 a month before shipping — the credibility note in the thread). Adoption note: pointer-CLAUDE.mds can be deleted; mixed repos need CLAUDE.md for Anthropic-specific layer since present CLAUDE.md suppresses AGENTS.md entirely. (3) Grok Build Remote Control in testing (via Musk RT) — the remote-session product filling in. (4) The eval layer self-organizes: Epoch AI launched Benchmark Reviews (15 audited: 4 verified, 9 flawed, 2 insufficient), WeirdML v3 (agentic, 11 hand-made tasks). (5) Ecosystem color: AgenticLinux (immutable agents-first desktop image), Jev on OpenRouter (System-One model: app state + typed query, no text generation).
2026-09-16 the funding layer keeps printing — Factory raised $200M (Blackstone, Khosla, Sequoia, NEA) for its "self-improving software development platform" (SiliconANGLE, Sep 15); and MIT CSAIL spun out G5 Labs ($14M seed, Pillar + Battery co-led, Jeff Dean angel; Kraska) for an "abstraction layer for AI-native software development" ("post-developer era") — dev-tools capital now assumes agents are the primary code-writers on both sides of the product. Galaxy Day 2 today; Day-1 color was testimonial-loop RTs (Katie Miller: had the bot unsubscribe her from recurring payments via email).

Claims to watch · 3 (3 pending, 0 slipped, 0 hit, 0 missed)

nichochar (X), Oct 3, 2026: Vercel "fundraise announcement incoming" review by Oct 10, 2026 overdue Source
Arav Srinivas (Perplexity), Oct 3, 2026: "Perplexity Computer on Vera" rollout review by Oct 17, 2026 overdue Source
nichochar, Oct 8, 2026: "lots of new agent protocols" in the next 6 months review by Apr 22, 2027 pending Source

All tracked claims

September 2026 · 3 updates

2026-09-11 ecosystem health, both sides of the ledger. OpenClaw v2026.9.4 (293 contributors; chats-to-skills, GPT Image 2.5 canvas, cloud controls) — fourth release in 11 days. Security posture published: 1,788 researcher reports ("most junk"), 17 critical, openclaw.ai/security. Companion tooling: Claude Code ships plugin eval (score plugins/skills with-vs-without per model release) and /diff as a persistent real-time pane — the harness layer building its own regression discipline. Grok Bot push: GTM integrations (Salesforce/HubSpot/Gong/Clay) + SpaceXAI livestream "build a company with Grok Bot" Sep 15–17 — spectacle-marketing, watch what actually ships.
2026-09-10 the harness layer gets a managed product: OpenAI's Agents API — "build and launch cloud agents," powered by the Codex harness, with orchestration, long-running sessions, tool use (posted 5:01 PM ET, unclassified at sweep). This is the OpenClaw/Grok Build/pi category becoming a vendor-bundled cloud service sold next to the model. Pricing/limits unknown; the open-ecosystem question is what "powered by the Codex harness" leaves for independent harnesses. Tracked in the new openai-bundling thread.
2026-09-07 Buzz itself is Block Inc.'s OSS human+agent workspace (Nostr relay = workspace; kind 40008 structured diffs; branch-as-room git hosting; YAML workflows) — not an OpenClaw feature. The @openclaw/buzz plugin IS OpenClaw-as-agent-in- Buzz (bot keypair, Bot role, up to 1,020 rooms/Gateway, room→agent bindings). Caution for adoption: Buzz's workflow approval gates are not wired end-to-end (runs hitting request_approval are marked Failed — WF-08; send_dm/set_channel_topic stubbed — WF-07; no rate limiting; unsigned alpha desktop builds). Buzz-as-forge is not yet a Forgejo replacement.

Background

The thread: the two harness-layer stories adjacent to the agent stack. OpenClaw 2.0 = v2026.8.1, shipped Aug 31, 03:30 UTC — a brand on the calendar version, not a separate line: 16,000+ PRs / 933 contributors (569 first-time), ~half of all PRs ever merged, after a ~7-week freeze (vs 106 releases in the prior 230 days). Headlines: guided onboarding (detects existing ChatGPT/Claude subs, API keys, local models), rebuilt browser app, shared cloud sessions (multiplayer), built-in memory + "grounded dreaming" consolidation + self-learning Skill Workshop, security overhaul (session permission modes, masked credential requests keeping secrets out of model context, plugin trust review), ~575ms Gateway start. Breaking: sessions→SQLite (back up first), OpenProse plugin removed, codex/→ openai/ — via openclaw doctor --fix. 2.0 also added first-class Buzz channels (relay conversations, message fidelity, typing indicators, NIP-27 mentions). v2026.9.2 (Sep 5) is the following calendar release on this line. Grok Build (xAI): terminal coding agent + TUI, teased Apr 2026, beta May 25, open-sourced Jul 15 (Apache 2.0, 26.3k stars, no external contributions), 1.0 Aug 7, now on Grok 4.6; Sep 7 update: first-party MCP server, long-running workspace daemon → Computer Hub, subagent continuation, background agents return output. Computer Hub is a hidden remote-workspace control plane gated by an xAI account flag (RuntimeWire, Aug 7) — foundation for an unannounced remote-session product. Category note: Grok Build is harness-class (pi/OpenClaw), NOT Buzz-class — Buzz is the messaging/relay layer.
Watch for: OpenClaw 2.0 follow-up patches (Gemini-embedding memory sync bug open as of Sep 1; the TTY doctor bug was fixed in-tracker), whether the two-month cadence holds, Grok Build Computer Hub activation (the xAI account flag), and whether xAI's marketplace converges on or forks from SKILL.md/AGENTS.md conventions.
Watch add: whether AA writes up the GDPval/Briefcase Grok runs (all our links are X-amplifications of AA runs); whether Fast's quality holds (same-model claim unverified); seat retention in Teslas.