# Agent Platform Research Briefing — July 24, 2026

Welcome to your agent platform research briefing for Thursday, July 24th, 2026.

**OpenAI launches Presence — enterprise agent deployment platform** — OpenAI dropped a major enterprise product on July 22nd called Presence, a platform for deploying trusted voice and chat AI agents in production. Presence includes built-in policy enforcement, guardrails, simulations, evaluations, approved actions, and Codex-powered agent testing. Analysts at Techi wrote a buyer's checklist on day one, framing it as OpenAI's move from raw model access to full-service agent consulting — "boots on the ground" pricing to manage enterprise deployments. This signals a strategic shift: OpenAI is no longer just selling tokens, it's selling turnkey operational AI. For agent developers, Presence could become the enterprise default deployment surface, similar to how Copilot dominated the coding agent space.

**Anthropic doubles AI lobbying spend to $40 million** — Anthropic announced July 21st an additional $20 million donation to Public First Action, bringing its total political spending to $40 million for the 2026 midterms — the largest single AI policy investment ever. The company is the leading voice for frontier AI regulation, pushing state-by-state progressive safety laws while Meta and OpenAI lobby against them. OpenAI's allied "Leading the Future" fund has a $125 million war chest. The donation cannot be used to elect specific candidates, but focuses on policy outcomes. This is the first election cycle where frontier AI labs are treating regulatory results as nine-figure bets. For the agent ecosystem, the regulatory landscape will shape what models are available, at what price, and under what constraints — making this a downstream factor in every agent architecture decision.

**AWS Kiro prompt injection vulnerability — agents rewriting their own configs** — AWS fixed a critical security chain in its Kiro coding agent where a poisoned web page could rewrite the agent's mcp.json configuration file and execute attacker-controlled code with full developer privileges — bypassing all approval gates. The exploit chain leveraged prompt injection to manipulate the agent into modifying its own MCP server configuration, essentially giving an attacker root access to the developer's environment. This follows a pattern established by Tenet Agentjacking in June and the Azure DevOps MCP hijack last week. The vulnerability is patched, but it underscores a growing reality: agent configuration files are the new attack surface. If your agent trusts MCP configs from unverified sources, you're vulnerable.

**Starship Flight 13 delayed again — weather forces July 24 attempt** — SpaceX pushed back Starship's 13th integrated test flight from Wednesday to today, Thursday July 24th, citing unfavorable weather conditions that would hamper heat shield imagery. This is the second delay — the original July 16th attempt was aborted at the last second when four of 33 Raptor 3 engines failed to ignite. Two engines were replaced between attempts. The mission's primary objective is capturing clear heat shield imagery during reentry, plus deploying the first Starlink V3 test satellites. If weather clears, launch is expected today around the same 6:45 PM CT window. After three attempts across ten days, the patience of both investors and engineers is being tested — and the FAA's closed Flight 12 investigation means any further issues would require a fresh review.

**GPT-5.6 becomes preferred model in Microsoft 365 Copilot** — OpenAI confirmed July 23rd that GPT-5.6 is now the preferred model powering Microsoft 365 Copilot across Word, Excel, PowerPoint, and Teams. This is the first major enterprise deployment upgrade since the model's public launch last week. For the agent ecosystem, it means millions of enterprise users are now interacting with GPT-5.6's enhanced reasoning, coding, and multimodal capabilities through the most widely deployed agentic interface on earth. The Microsoft integration validates GPT-5.6 as the enterprise default and puts pressure on competing coding agents to match or exceed Copilot's capabilities at comparable scale.

**Claude Opus 5 NOT released — July 23 target missed** — The prediction markets and social media buzz pointed to Thursday July 23rd as Anthropic's launch date for Claude Opus 5, codenamed Honeycomb. It didn't happen. Opus 4.8 remains the public frontier flagship. The Honeycomb EAP model briefly appeared in Cursor's model picker on July 9th and was spotted on Google Vertex AI's Model Garden on July 14th, but neither has been confirmed by Anthropic. Prediction markets showed strong confidence for a late-July Opus 5 launch earlier this week, but the miss on the consensus Thursday date is notable. Meanwhile Anthropic shipped tooling improvements instead — ClaudeDevs' chartography cookbook saw Fable 5 go from 29% to 73% success rates, and Claude Voice was upgraded to run Opus/Sonnet with connectors. Expect an announcement by end of July, but don't hold your breath today.

That's the briefing for today. Stay sharp out there.