On Sat Aug 8, Anthropic's Claude Code ships two patches inside one day — v2.1.225 lands gateway spend-limit surfacing (the limit-reached message now names the cap, its reset time, and the operator's message), a workspace-trust prompt on the claude agents surface for untrusted directories, SendMessage able to open a fresh conversation with a named Remote Control session on another machine (previously you could only reply after they messaged first), fixes for an MCP OAuth macOS keychain-timeout that produced burst 401 errors as if never authenticated, an auto-mode consecutive-block counter that had wrongly counted the model's own safety-filter refusals, a Remote Control conversation-history breakage on resume after very-large-conversation compaction, a headless-session issue where cross-session messages stayed parked with no notice or expiry, a hovering-over-another-project's-session-in-the-agents-list changing the next agent's working directory, a Claude Code on the web false-stuck signal that re-sent a growing event backlog on every reconnect, and a claude self-hosted-runner registration-then-fail loop when --base-dir cannot be created or written; v2.1.226 follows the same afternoon as a bug-fixes-and-reliability patch on top. On Fri Aug 7, OpenAI's Codex keeps its sub-daily alpha cadence with rust-v0.148.0-alpha.4 and alpha.5 stacked on the Aug 7 v0.147.0 stable cut (prior edition), and Google Antigravity ships 2.6.0 with the antigravity_guide builtin skill (instant in-context reference for Antigravity 2.0, CLI, IDE, and SDK), syntax highlighting for C++, Python and Protobuf in markdown code and diff blocks, a centralised Multi-Model Configuration that replaces the legacy singular gemini_config with a unified repeated models collection on AgentConfig and LocalAgentConfig for multi-model routing and fallback, and Enterprise sign-in support for Gemini Enterprise accounts and Workforce Identity Federation via Advanced SSO. On the consumer side, OpenAI on Thu Aug 6 removes text-chat limits for Free and Go users and makes GPT-5.6 Luna the default for those tiers (replacing GPT-5.5 Instant), rolls out a Think button that lets Free and Go users toggle extra reasoning, and adds a manual compute-effort slider for Plus and Pro users — the rate limits (10–40 messages per 3–5 hours on flagship models) that had previously kicked users down to a lighter model are gone for text chat; file uploads, image attachments, image generation, voice mode and other resource-intensive features are still rate-limited. On the agent-perimeter capital tape, Obsidian Security on Tue Aug 4 closes $85M Series D led by Crescent Cove Advisors at a $1.1B valuation (unicorn) for its non-human-identity and AI-agent security platform, and extends native governance controls to Anthropic's Claude Code and Cowork to let security teams restrict agent permissions around production data, manage access to sensitive files, and block unsanctioned MCP or tool usage at runtime; OLIX on Mon Aug 3 closes $312M Series B led by Fundomo at a $3.3B valuation (up from ~$1B at February's $220M Series A) with Arm, Hudson River Trading, Reed Hastings and the UK Sovereign AI venture fund for its photonic DX-1 decode accelerator (SRAM instead of HBM) targeting first customers in H2 2027, and appoints Stanford's Nick McKeown to its board. On the agent-plumbing GitHub tape, four fresh repos land inside the same seven days: alikon-art/DeterminFlow (Sun Aug 2) hits 271 stars as a production-oriented Python AI-workflow runtime for building, validating, recovering and shipping complex AI workflows as dependable services with a bilingual English + Chinese README; i3T4AN/KADATH (Sat Aug 8) hits 151 stars in hours as an evolutionary multi-agent runtime that breeds, evaluates and improves autonomous agents across reproducible epochs; oliverb-io1902e8/agent-skills-collection (Thu Aug 6) hits 146 stars as a curated collection of modular agent skills for LLM-based agents; and PatilShreyas/debroid (Sun Aug 2) hits 98 stars as an autonomous headless Android debugger that lets AI coding agents inspect runtime memory, set breakpoints and debug live apps, tagged for Claude Code, Codex, and Google Antigravity.
— the throughline is a perimeter week: the Claude Code, Codex and Antigravity harnesses all tighten the operator-controls surface (gateway spend-limits, workspace trust, Enterprise SSO, alpha-cadence patches), OpenAI democratises the free-tier chat surface and puts the flagship reasoning toggle in the Free-tier UI, and the agent-perimeter capital tape lands ~$400M inside 48 hours across Obsidian's non-human-identity security and OLIX's inference-chip perimeter.

Two reads. (1) The gateway spend-limit surfacing is the operator-facing half of the enterprise-gateway story. The gateway (Anthropic's enterprise proxy that governs which requests reach the model and under what spend cap) has been enforcing spend caps for months; what was missing was the shape a developer sees when the cap is hit: a bare rate-limit error. v2.1.225 makes the limit-reached message carry the cap name, its reset time, and the operator's custom message, which is the shape a compliance surface takes when the buyer wants the developer to know why they were stopped and who to contact. Read alongside v2.1.223's owner/* marketplace wildcards (prior edition) and v2.1.224's archive plugin source with SHA-256 pinning (prior edition), the Aug 2026 Claude Code posture is the enterprise policy layer now writes plainly to the developer. (2) SendMessage-by-name to a remote machine is the practical continuation of v2.1.224's cross-machine SendMessage: previously you could only reply to a machine that had messaged you first; now you can open a conversation with a machine by name. That is the shape a fleet primitive takes when the orchestrator wants to reach specific runners, not only respond to inbound.

Two reads. (1) Two-cuts-same-day is the “we found a regression in v2.1.225 and we shipped the fix inside hours” signal, not a scheduled cadence bump. When the marquee release is v2.1.224's self-hosted runners (prior edition) — a fleet primitive that changes the shape of Team- and Enterprise-plan deployment — the tail of fixes is where the real production surface gets caught. v2.1.225's ten fixes plus v2.1.226's follow-on are the shape a coding-agent harness takes when the buyer is running v2.1.224 in production before v2.1.224 has had a full day to soak. (2) Three cuts in three days is the loudest same-week cadence Claude Code has shipped in 2026. Read alongside Codex's v0.148.0-alpha.4 and alpha.5 in the same 24 hours (item 03) and Google Antigravity 2.6.0 on Fri Aug 7 (item 05), the Aug 7-8 coding-agent-harness shape is all three reference harnesses ship inside the same week.

Two reads. (1) The alpha cadence itself is the story. Fri Aug 7's v0.147.0 stable (prior edition) shipped portable Agent Plugins, MCP 2026-07-28 opt-in and --approve-for-me; within 24 hours, the v0.148 alpha train is already at alpha.5. That is the same posture Codex has run through v0.143 → v0.147: stable cuts release features, alpha trains stack pre-releases at sub-daily cadence, the next stable inherits the top of the alpha stack. (2) Codex's sub-daily alpha cadence and Claude Code's three-cuts-in-three-days v2.1.224 → v2.1.226 (items 01, 02) are the two shapes of the coding-agent-harness ship train under load: Codex stacks alphas continuously off a stable base; Claude Code ships fewer alphas but tighter same-week clusters when a marquee lands. Read alongside Google Antigravity 2.6.0 on Fri Aug 7 (item 05), the Aug 2026 shape is all three reference harnesses are on aggressive weekly-or-tighter cadence at the same time.

Two reads. (1) The text-chat limit removal is the loudest consumer-side signal of the summer and a shape shift in what “free tier” means for a frontier LLM. Previously, ChatGPT Free was rate-limited to 10–40 flagship messages in a 3–5 hour window, then downshifted to a lighter model; now the default itself is the lighter model (GPT-5.6 Luna) and text chat is unlimited. That is the shape a frontier lab takes when Luna's per-token cost has fallen far enough that unlimited free chat is cheaper than the acquisition value of the ceiling. Read alongside OpenAI's Jul 30 GPT-5.6 Luna 80% price cut and Terra 20% cut (prior edition), the Aug 2026 posture is Luna is priced to be the free default. (2) The Think button for Free / Go is the interesting UI concession: the reasoning toggle that Pro users used to pay for is now a button in the free UI. That is the shape a consumer AI product takes when the marginal cost of a per-query reasoning burst is less than the risk of the free-tier user perceiving the product as “dumber”.

Two reads. (1) The Multi-Model Configuration is the story. Antigravity's legacy config assumed one gemini_config per agent; 2.6.0 replaces that singular field with a repeated models collection on AgentConfig and LocalAgentConfig supporting multi-model routing and fallback. That is the shape a coding-agent harness takes when the buyer wants to route different sub-turns to different models (Gemini for one leg, an external model for another, a smaller Gemini for cheap edits) — and the shape a Google-owned harness takes when it concedes that not every sub-turn should be Gemini. (2) The Enterprise SSO cut — Gemini Enterprise + Workforce Identity Federation via Advanced SSO — is the compliance surface that enterprise-plan Antigravity was missing. Read alongside Claude Code v2.1.224's self-hosted runners on Team + Enterprise (prior edition), the Aug 2026 harness shape is each of the three reference harnesses ships an enterprise-tier primitive in the same 48 hours.

Two reads. (1) The Claude Code and Cowork native-governance extension is the operative product-milestone half of the raise. Non-human identity as a category has been growing since service accounts became a first-class citizen in cloud IAM; the AI agent is the newest and most complex non-human identity. Obsidian's extension means a security team can restrict what an agent can touch in production, which files it can read, and which MCP servers or tools it can invoke at runtime, from a single control plane. That is the shape a non-human-identity security product takes when Claude Code and Cowork are the buyer's reference agent surfaces. Read alongside Claude Code v2.1.223's owner/* marketplace wildcards (prior edition) and v2.1.225's gateway spend-limit surfacing (item 01), the Aug 2026 agent-perimeter shape is enterprise policy lands on the buyer's security control plane, and the harness surfaces the enforcement to the developer. (2) The capital-signal is 60 of the Fortune 500 already on the platform: this is the round funded off distribution, not product proof. Crescent Cove writes growth checks; the $85M at $1.1B is the shape a Series D takes when the account base is already the reference.

Two reads. (1) The DX-1 architecture — SRAM instead of HBM — is the technical thesis. HBM is the current inference bottleneck (HBM supply is short, packaging is expensive, and the bandwidth is the ceiling on decode throughput). OLIX's bet is that a chip built on on-die SRAM plus optical networking can skip advanced packaging entirely and ship a decode accelerator without HBM dependence. That is the shape a frontier inference chip takes when the buyer is willing to trade absolute-model-size ceiling for supply-chain independence and lower latency-per-decoded-token. Read alongside Anthropic's Wed Aug 5 in-house silicon team (prior edition), the Aug 2026 inference-chip capital tape is hyperscalers and labs are pursuing custom silicon in parallel because the HBM+packaging bottleneck is the shared ceiling. (2) The Reed Hastings and UK Sovereign AI Fund participation is the political-capital signal: a US strategic angel plus a UK sovereign-wealth participation is the shape a UK chip Series B takes when the UK government has decided that sovereign inference capacity is strategic infrastructure. Prof. Nick McKeown on the board is the deep-technical signal: OpenFlow / SDN is the reference programmable-networking architecture, and OLIX's optical networking is the same discipline applied at chip scale.

Two reads. (1) DeterminFlow is the shape the “durable AI workflow” category takes when the author decides Temporal + LangGraph is too heavy and n8n too UI-first and ships a Python-first runtime with agent-orchestration as a first-class primitive. 271 stars in six days is the shape a real gap in the workflow-runtime slot takes when the tool that fills it is FastAPI-native and can be dropped straight into a Python service already exposing HTTP endpoints. (2) The bilingual English + Chinese README is the signal that matters: a Chinese-team-authored open-source runtime that keeps English as a first-class documentation surface is the shape a Chinese-open-weight-ecosystem-adjacent tool takes when the author explicitly wants Western adoption. Read alongside AMAP-ML's LongHorizon-Harness (prior edition) and Anionex's agent-vision-toolkit (prior edition), the Aug 2026 shape of the Chinese-authored AI-agent-plumbing cohort is the plumbing library ships with bilingual docs and native support for both Chinese and Western harnesses.

Two reads. (1) Evolutionary multi-agent runtimes are the interesting shape the “optimise the agent itself, not the prompt” discipline takes when the author reaches for genetic-algorithm primitives. Most prompt-optimisation libraries optimise a single prompt against an eval; KADATH optimises a population of agents against a goal, letting fitter variants survive successive epochs. That is the shape a self-improving agent framework takes when the author treats the agent (system prompt + tool set + orchestration policy) as the unit of evolution. (2) 151 stars within hours of creation on Aug 8 is the shape a genuinely novel primitive takes when the AI-agent-plumbing OSS crowd is already looking for it. Read alongside AMAP-ML's long-horizon harness (prior edition) and oliverb-io1902e8's agent-skills-collection (item 10), the Aug 2026 shape of the agent-plumbing cohort is every week ships a new primitive that the ecosystem has been implicitly waiting for.

Two reads. (1) The “modular agent skills” primitive is consolidating across harnesses at pace. Anthropic's SKILL.md format is the reference; Codex v0.147.0's Cursor-managed skill imports (prior edition) is the cross-harness bridge; agent-skills-collection is the shape a third-party skill-library takes when the author decides to package skills once and target multiple harnesses. Read alongside 0xwilliamortiz/claude-red (prior edition), the Aug 2026 shape of the skills-library category is curated skill packs, tagged for both Claude and Codex, growing to 100+ stars inside 48-72 hours. (2) The “curated” framing itself is the differentiator: the noise floor on SKILL.md dumps is already high; curated collections that the author has vetted and organised are the shape the skills-library market takes when quality wins over quantity.

Two reads. (1) Mobile-app runtime debugging is a workflow gap the coding-agent harnesses have not yet plugged: Claude Code, Codex and Antigravity all run terminal, file-write and web-browse tools well but none of them can attach to a live Android app process, set breakpoints, and inspect runtime memory autonomously. debroid fills that slot as a Kotlin CLI the agent can invoke — and the shape a coding-agent workflow takes when the agent needs to debug a running Android app, not just write code for it. (2) The antigravity topic tag is the notable inclusion: most agent-plumbing repos tag claude-code and codex only; debroid explicitly targets Google Antigravity as a first-class harness. Read alongside Google Antigravity 2.6.0's antigravity_guide builtin skill (item 05), the Aug 2026 shape of the harness-plurality signal is the third-party plumbing cohort is treating Antigravity as a tier-one harness alongside Claude Code and Codex, not an also-ran.