On Fri Aug 7, the two reference coding-agent harnesses ship major cuts on the same day — Anthropic's Claude Code v2.1.224 lands self-hosted runners for Team + Enterprise, cross-machine SendMessage and ListAgents so sessions can talk across your fleet, JWT-aware credential masking, AWS SigV4 re-signing, an archive plugin source with SHA-256 pinning, and removes the 200-subagent-per-session spawn cap; OpenAI's Codex v0.147.0 stable ships portable Agent Plugins with local / personal / workspace / remote catalog search, an --approve-for-me flag for auto-reviewed approvals, opt-in support for the MCP 2026-07-28 protocol (paginated discovery, multi-round requests, non-blocking server startup), Cursor-managed skill imports, bearer-token redaction across displayed commands and replay, and removes the deprecated codex exec --full-auto flag. On Thu Aug 6, Claude Code v2.1.223 lands four permission-bypass patches at once: a Bash permission bypass where a crafted command could hide parts of itself from permission checks; a permission-prompt bypass using tabs and invisible Unicode to hide part of the command from the approval dialog; a workflow-script sandbox escape using dynamic import() to run code outside the sandbox; and a permission gap where an agent definition's bypassPermissions mode ignored the org-level bypass-permissions disable policy. On the agentic-CX capital tape, Klaviyo (NYSE: KVYO) agrees on Wed Aug 5 to acquire the 25-person Agency team and technology and install co-founder Elias Torres as Chief Product Officer post-close (Q3 2026) to lead the Composer campaign-builder and Customer Agent post-sale-support product lines, reporting to co-CEO Andrew Bialecki; Omilia on Thu Aug 6 closes a $67M Series B led by Expedition Growth Capital for its agentic self-learning voice-first CX platform already running at Capital One, Discover and Taco Bell, on live ARR up 10× since Series A to over $60M and no interim equity raise; and Naïve on Wed Aug 6 lands a $28.5M Series A led by Nexus Venture Partners (with Y Combinator, Zetta, Liquid 2, JD Sherman, Gokul Rajaram, Zachary Sims, Robert Chatwani) for its unified API + serverless stack for autonomous companies, on 30,000+ developer signups within months and ARR up 10× to low-double-digit millions in six months. On the agent-plumbing GitHub tape, petergyang/human-review holds 574 stars as a visual Google-Doc-style comment layer that lets a reviewer mark up HTML and Markdown produced by an AI agent and push feedback back into the harness; 0xwilliamortiz/claude-red ships on Wed Aug 5 and hits 559 stars in three days as a curated library of offensive-security SKILL.md files for Claude and Codex (SQLi, shellcode, EDR evasion, exploit development); cristicretu/diri ships on Tue Aug 4 as a native macOS Rust / Swift / GPUI orchestrator that runs Claude Code, Codex, Cursor, Gemini and shells in parallel across git worktrees and remote hosts (222 stars); AMAP-ML/LongHorizon-Harness ships on Tue Aug 4 as Alibaba AMAP-ML's Python long-horizon computer-use harness with fresh-context execution, durable verified state, independent auditing, recoverable progress, and native Claude Code / Codex / OpenClaw integration (387 stars); and Anionex/agent-vision-toolkit ships on Fri Aug 1 as a vision toolkit and skill for text-only LLMs (DeepSeek, GLM, Kimi) covering image Q&A, long-screenshot OCR, frontend UI restoration and GUI automation with seamless integration for Codex, Claude Code, Pi, Oh My Pi and OpenCode (348 stars).
— the throughline is a harness week: on the same 48 hours the two reference coding-agent harnesses both tighten the permission model and extend the fleet primitive (self-hosted runners, cross-machine SendMessage, MCP 2026-07-28 paginated discovery, --approve-for-me auto-approval), the agentic-CX buyer lands ~$150M of Series capital across three product lines (Composer + Customer Agent, voice, autonomous-company infrastructure), and five agent-plumbing repos cross the 200-star line filling the skills, review, orchestration and vision-for-text-only-models gaps around the harnesses.

Two reads. (1) Self-hosted runners is the fleet primitive Claude Code has been reaching for since cross-session SendMessage arrived in the v2.1.16x cycle in July (prior editions). Cross-session was same-machine; v2.1.224 makes it cross-machine, and self-hosted-runner makes your own hardware the hosting substrate for web, mobile and desktop sessions. That is the shape a coding-agent harness takes when the enterprise buyer will not accept Anthropic-run runtime for data-sovereignty reasons — the self-hosted lane is the compliance answer, and cross-machine SendMessage is the collaboration answer on top of it. (2) The removal of the 200-subagent-per-session spawn cap is the “we-tested-and-the-workflow-hits-the-cap” signal: real production workflows are routinely spawning more than 200 subagents in a single session, which is the shape a long-horizon-harness workload takes when the parent orchestrator is doing multi-repo migrations, PR-fleet reviews or codebase audits. Read alongside Codex v0.147.0's Agent Plugins (item 03) and the MCP 2026-07-28 stateless spec (prior edition), the Aug 2026 shape is the harness is now the fleet manager, not the terminal.

Two reads. (1) The four patches are the four different shapes a permission bypass can take in a coding-agent harness with a live permission model: command-string obfuscation (Bash), rendering-layer obfuscation (tabs and invisible Unicode against the approval dialog), runtime escape (dynamic import() out of the workflow sandbox), and policy-layer bypass (agent-level bypassPermissions overriding org policy). Landing all four in one cut is the shape a security review by real customers takes when it reports upward across a category, not one finding at a time. (2) The owner/* wildcards for strictKnownMarketplaces and blockedMarketplaces are the “we-have-to-govern-marketplace-orgs-not-individual-repos” primitive: enterprise policy now writes “allow all plugins under owner X, block all plugins under owner Y”, which is the shape a plugin marketplace takes when the buyer is an enterprise security team, not an individual developer. Read alongside v2.1.224's archive plugin source with optional SHA-256 pinning (item 01), the Aug 6-7 Claude Code shape is lock the plugin supply chain and lock the permission surface, in the same 24 hours.

Two reads. (1) Agent Plugins with a four-tier catalog (local + personal + workspace + remote) is the shape a coding-agent harness takes when it accepts that the plugin surface is now four different governance surfaces: the developer's own machine, their personal plugin bundle, their team workspace, and the shared remote catalog. That is the same four surfaces that Claude Code v2.1.223's owner-wildcard marketplace policy is solving on the Anthropic side (item 02). (2) Opt-in for the MCP 2026-07-28 protocol is the concrete first-mover adoption of the stateless MCP spec that landed Jul 28 (prior edition): Codex now speaks paginated discovery, multi-round requests, and non-blocking server startup to any MCP server that supports the 2026-07-28 wire format. The --approve-for-me flag is the “we-know-most-approvals-are-noise” primitive: Codex reviews the approval itself against the sandbox and workspace policy and gates only the ones that need a human. Read alongside Claude Code v2.1.224's cross-machine SendMessage and self-hosted runners (item 01), the Aug 7 coding-agent-harness shape is the two reference harnesses ship the same fleet + permission primitives on the same day, against the same MCP wire.

Two reads. (1) The operative read is Klaviyo is buying the CPO seat, not the technology. Composer and Customer Agent already ship inside Klaviyo; Agency's 25-person AI-native team is the accelerant; and Torres — a Drift co-founder who previously built the go-to-market for a category-defining conversational product — is the profile a publicly-traded B2C CRM reaches for when it wants to convert “agents that build marketing campaigns and answer post-sale questions” into a defensible product line. That is the shape a market-leader CRM takes when the agent surface is now the primary product surface. (2) The Sequoia / Menlo / Felicis exit into a public-market strategic is the shape a venture-backed AI-native GTM co takes when the market has already picked its incumbent: the technology sells for team-and-CPO-seat, not independent trajectory. Read alongside Omilia's $67M Series B on $60M ARR (item 05) and Naïve's $28.5M Series A on 30K devs (item 06), the Aug 5-6 agentic-CX capital shape is strategic acquires the GTM CPO, independents raise growth.

Two reads. (1) The 10× ARR without an intervening equity round is the profile a growth-equity buyer reaches for: Expedition Growth Capital is writing at $60M+ ARR with Capital One, Discover and Taco Bell already live, which is the shape an enterprise-agentic-CX buyer takes when the platform is already the reference and the round funds distribution, not product. (2) Voice-first, self-learning is the differentiator: most agentic-CX competitors are chat-first with voice bolted on; Omilia is voice-native. Read alongside xAI's Grok Voice Think Fast 2.0 GA in July (prior editions), Sesame's $250M Series C (prior editions), and Rime's $24M for enterprise speech-to-speech (prior editions), the Aug 2026 shape of the voice-agent segment is the platform layer (Omilia) is raising off enterprise ARR while the model layer (Grok Voice, Sesame, Rime) is raising off benchmark leadership.

Two reads. (1) The customer roster is the story. AI automation agencies, faceless TikTok / YouTube channels, and one rental-car agency are the buyer profile of a service that automates the grunt work of setting up and running a company: one-founder or zero-founder shops that need the legal, financial and cloud spine of a real company but cannot afford, and do not want, an accountant and a lawyer and a devops engineer. That is the shape a bottom-up market takes when the incremental cost of “spin up a real corporate entity that runs itself” falls below the cost of the person who would have done it. (2) The 6-month 10× ARR and 30K+ dev signups in months are the traction signals a Nexus + Y Combinator + Zetta round is writing off; the angel bench (Rajaram, Zheng, Sherman, Chatwani, Sims) is the operator-adjacency signal. Read alongside Omilia's $67M enterprise voice round (item 05) and Klaviyo's Agency-and-Torres deal (item 04), the Aug 5-6 agentic-CX capital tape spans the top of the market (Klaviyo strategic), the enterprise segment (Omilia) and the long-tail one-founder segment (Naïve) — each with a different capital structure.

Two reads. (1) Human-review is the missing HITL primitive between raw agent output and PR-review UI. A coding agent produces HTML or a Markdown report; the reviewer wants to comment on line 42 without opening a PR. Human-review is the shape that gap takes when someone fills it as a single-purpose tool, and 574 stars in ten days is the shape a real gap in the workflow takes when the tool that fills it lands well. (2) The topic tags spell out the intended surface — ai-agents, claude-code, codex, human-in-the-loop — and the harness-agnostic README is the “works inside your favourite AI harness” commitment. Read alongside Claude Code v2.1.224's cross-machine SendMessage (item 01) and Codex v0.147.0's Agent Plugins (item 03), the Aug 2026 shape is the tooling around the harnesses is filling in faster than the harnesses can ship the primitives themselves.

Two reads. (1) The skills-as-a-library shape has already crossed the offensive-security threshold: claude-red is the second high-star offensive-skills pack of the summer after the offensive-claude pack of prior editions, and the 559-stars-in-three-days trajectory is the shape a skill library takes when the pen-testing community is tooling on top of Claude Code the same way it once tooled on top of Metasploit. That is a real product-market fit signal for the Claude skills primitive, distinct from the general marketplace. (2) The topics tag is the harness-agnosticism signal: claude + claude-code + claude-skills + codex + codex-skill + skill + skills means the same SKILL.md format now travels across two harnesses, which is exactly the surface Codex v0.147.0's Cursor skill import addresses on the other side (item 03). Read alongside Claude Code v2.1.223's owner/* marketplace wildcards (item 02) and v2.1.224's archive plugin source with SHA-256 pinning (item 01), the Aug 2026 skills-supply-chain shape is rich unregulated library, tightening org-level policy, portable across harnesses.

Two reads. (1) Diri is the native-macOS shape of the operator-console category: run multiple coding agents in parallel, each on its own git worktree, each optionally on a remote host. The Electron cohort of the summer (BossConsole, Stably Orca, prior editions) hit the same primitive from a cross-platform direction; diri is the platform-native answer, keyed to Rust + Swift + GPUI. That is the shape a coding-agent-first developer tool takes when native performance and macOS integration are the differentiator. (2) Native git-worktree parallelism is the practical unlock: an operator can fan out four coding sessions, each on a different branch, each editing the same repo without conflicts, and merge back through the diri UI. Read alongside Claude Code v2.1.224's cross-machine SendMessage (item 01) and Codex v0.147.0's Agent Plugins (item 03), the Aug 2026 operator-console shape is the harnesses ship fleet + skills primitives, and the operator-console layer ships the multi-harness parallel UI on top.

Two reads. (1) Alibaba AMAP shipping a public harness is the story: AMAP is China's reference mapping-and-navigation product, its ML org is production-scale, and a MIT-licensed long-horizon harness from that org is the shape a hyperscale Chinese team takes when it wants Western coding-agent harnesses to run inside its research workflows. Fresh-context execution + durable verified state + independent auditing + recoverable progress is the four-part “keep the agent honest across many hours” primitive, which is the shape a long-horizon-harness workload needs when the base agent context window is the ceiling. (2) The OpenClaw integration is the notable third-party: OpenClaw is the open-source coding-agent runtime the Chinese open-weight ecosystem has adopted as its harness equivalent, and native support alongside Claude Code and Codex is the shape a plumbing library takes when it wants to be the neutral harness-independent layer. Read alongside Anionex's agent-vision-toolkit (item 11), the Aug 2026 Chinese-open-weight tooling shape is ship the plumbing on top of Claude Code, Codex and OpenClaw at the same time.

Two reads. (1) Agent-vision-toolkit is the shape the “Chinese open-weight text model + Western vision + Western harness” stack takes when someone stitches it together as a single skill. DeepSeek, GLM and Kimi are the three reference Chinese open-weight text models; none of them ships a first-party vision head at frontier parity; the toolkit routes image inputs through an external vision-language model and returns text the DeepSeek / GLM / Kimi caller can act on. That is the shape a multimodal-gap workaround takes when the developer chooses the text model for cost or licensing and needs vision on top. (2) The integration matrix — Codex, Claude Code, Pi, Oh My Pi, OpenCode — is the “every harness the buyer might actually run” commitment. Read alongside AMAP-ML's LongHorizon-Harness (item 10), the Aug 2026 shape of the Chinese-open-weight-plus-Western-harness pattern is the plumbing library is now the seam: Chinese labs ship text weights, Chinese and Western developers ship the skills that make those weights usable inside Claude Code and Codex.