OpenAI closed May with its most consequential week of the year on multiple fronts. Workspace Agents reached general availability across Business, Enterprise, and Edu. The Agents SDK got a fundamental architectural rethink separating harness logic from compute. Goal Mode is now GA with full admin analytics. OpenAI filed a confidential S-1 for a Q4 2026 IPO. A new $4B Deployment Company JV moves OpenAI into enterprise services. And Codex CLI v0.135.0 shipped with the most substantial diagnostic and editing improvements in months. This is the Codex Weekly for May 29, 2026.

The Biggest Story: A New Agents SDK Architecture

The week’s most consequential development didn’t ship as a changelog entry β€” it came from a May 28 Build Hour session where OpenAI introduced a model-native harness as the next design pattern for the Agents SDK.

The core shift: the harness (control loop, snapshotting, rehydration, tool orchestration) is now explicitly separated from the compute layer (the sandboxed containers or VMs where code actually runs). Previously these concerns were tangled together, which made it hard to recover from failures, swap sandbox providers, or enforce isolation boundaries without rebuilding core agent logic.

The new model formalizes four first-class primitives:

  • MCP (stdio and HTTP) for connecting external tools and data sources
  • Skills β€” versioned, reusable capabilities stored in GitHub (think “run tests” or “deploy service”)
  • AGENTS.md β€” a declarative manifest for agent capabilities with external storage mounts (S3, R2, Azure Blob, GitHub)
  • shell + apply-patch for file-system-aware code editing loops

The Responses API replaces the Assistants API checkpointer model, with runtime state persisted via a response_id chain. Supported pluggable sandbox providers at launch: E2B, Modal, Cloudflare, Vercel, Daytona, Blaxel, and Docker.

For enterprise teams that have been waiting for a stable SDK shape before committing to production agentic builds, this is the architecture worth standardizing on. Session code lives in the openai/build-hours GitHub repo; the full video is on YouTube. (StartupHub.ai)

For a deeper look at how this compares to Claude Code and Gemini CLI harnesses in an enterprise context, see our enterprise guide to agentic coding harnesses.

Workspace Agents GA + Enterprise Admin Wave

ChatGPT Workspace Agents hit GA on May 21 across Business, Enterprise, and Edu. Agents can own full workflows autonomously, follow team processes, and be shared across the workspace directory β€” build once, deploy to everyone.

May 28 added a wave of admin controls that close several gaps that had kept enterprise rollouts cautious:

  • Role-based publishing permissions β€” admins can lock down which roles publish agents to the workspace directory. This is the governance control that enterprise security teams have been asking for.
  • GPT-5.5 with configurable reasoning effort β€” set the model and effort level at build time rather than per-session.
  • Audio file output β€” agents can now produce audio as part of a response.
  • Smarter Slack thread replies β€” connected agents respond to relevant follow-up messages in a thread, with creator-controlled trigger behavior.

App Templates for GitHub Enterprise, Snowflake, and Databricks are now available to Enterprise and Edu admins. Each provides a guided setup flow covering OAuth credentials, webhook details, managed MCP server URLs, and workspace access controls before publishing. For organizations already running these platforms, this collapses what was previously a multi-hour integration exercise into a guided wizard. (OpenAI Enterprise Release Notes)

Goal Mode GA, Admin Analytics, and Skills Governance

The May 21 enterprise release also delivered three capabilities worth calling out separately.

Goal Mode is generally available in the Codex app, IDE extension, and CLI. Users define an outcome and success criteria; Codex works toward it autonomously. This closes the loop between task-level and outcome-level delegation β€” Codex isn’t just completing steps, it’s evaluating whether the goal has been reached.

Global admin console analytics for Codex now cover active users, credits/tokens consumed, threads/turns, user leaderboards, plugin usage, accepted lines of code, and model usage. For team leads trying to build a business case around Codex adoption, this is the data layer that makes it possible.

Skills governance shipped with the controls enterprise teams need to feel comfortable: skills are off by default in Enterprise/Edu workspaces, uploaded skills are scanned before activation (borderline flagged, risky blocked), and skill_id appears in Compliance Log conversation event streams for full traceability. (OpenAI Enterprise Release Notes)

The Partnership Momentum

Three enterprise deals landed this week that reflect how broadly Codex is being deployed:

Cisco (May 27) deepened its OpenAI partnership to deploy AI-native development across its entire engineering organization, including defect remediation and AI Defense work. This isn’t a pilot β€” Cisco is treating Codex as infrastructure for its dev org. (OpenAI)

Dell Technologies (May 18) partnered with OpenAI to bring Codex into hybrid and on-premises environments via the Dell AI Data Platform and Dell AI Factory. For enterprises that can’t route code through cloud sandboxes due to data residency or security requirements, this is the on-ramp. (OpenAI)

Gartner recognized OpenAI as a Leader in the 2026 Magic Quadrant for agentic coding (May 22) β€” the kind of validation that accelerates procurement approvals in large organizations. (OpenAI)

Also noteworthy: OpenAI published two case studies this week. Virgin Atlantic used Codex to achieve near-total unit test coverage and ship a revamped mobile app on a fixed holiday deadline. Endava built an “agentic organization” with Codex, cutting requirements analysis from weeks to hours. The velocity signal from real enterprise deployments is consistent.

OpenAI Company & Strategic Moves

IPO Filing

OpenAI filed a confidential draft S-1 registration statement with the SEC on May 22, signaling a Q4 2026 IPO (estimated September–November window). The company has engaged Citigroup and JPMorgan as advisors. Analysts estimate a target valuation of $852B–$1T and roughly $60B in new capital to be raised β€” against a projected ~$14B net loss in 2026 on $13.1B in 2025 revenue. For enterprise IT buyers, a public listing would bring quarterly financial transparency that currently doesn’t exist, changing the long-term vendor commitment calculus. (IndMoney)

OpenAI Deployment Company: $4B Enterprise JV

The most strategically significant announcement for enterprise buyers: OpenAI launched the OpenAI Deployment Company (DeployCo), a $4B+ joint venture to embed “forward deployed” AI engineers directly into client organizations. The JV is led by TPG and co-led by Advent International, Bain Capital, and Brookfield, with BBVA, Goldman Sachs, McKinsey, and Capgemini as founding partners.

This moves OpenAI downstream from API and platform provider into end-to-end enterprise transformation services β€” a direct step onto territory held by Accenture, Capgemini, and Big 4 consultancies. For organizations evaluating AI implementation partners, DeployCo represents a new, well-capitalized option with direct access to OpenAI’s model and product roadmap. (Cleary Gottlieb)

Anthropic Eclipses OpenAI in Valuation

On May 28, Anthropic raised $65 billion in a Series H round, pushing its valuation to $965 billion β€” surpassing OpenAI’s last private valuation for the first time. The round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, each investing more than $2 billion. For enterprise teams tracking the competitive landscape, investor confidence in Anthropic is accelerating in parallel with OpenAI’s IPO preparation. (Bloomberg)

Frontier Governance Framework

OpenAI published its Frontier Governance Framework (May 28) β€” a public document explicitly mapping internal safety practices to California’s SB 53 (Transparency in Frontier AI Act) and the EU AI Act’s code of practice for general-purpose AI systems. Coverage areas include cyber offense, CBRN threats, harmful manipulation, and loss of control.

For compliance and legal teams, this is OpenAI’s first formal, public regulatory mapping β€” relevant if your organization is assessing AI vendor governance obligations ahead of EU AI Act enforcement timelines. (OpenAI)

Rosalind Biodefense

OpenAI launched Rosalind Biodefense (May 29), expanding trusted access to GPT-Rosalind β€” a frontier life-sciences reasoning model β€” for vetted developers and U.S. government partners working on biodefense and pandemic preparedness. Access is gated through a tiered vetting process for researchers, government agencies, and trusted institutions. (OpenAI)

Election Safeguards 2026

OpenAI detailed its election safeguards strategy (May 27): live vote count integrations with the Associated Press, cybersecurity support for election administrators, and multi-layer content provenance measures for deepfakes. Aligns with broader industry watermarking momentum β€” SynthID standards are being adopted by OpenAI, NVIDIA, and ElevenLabs ahead of EU Article 50(2) provider-side requirements. (OpenAI)

Media Partnerships

OpenAI announced a strategic content partnership with Grupo Folha (Folha de S.Paulo) and Grupo UOL (May 25), bringing trusted Brazilian journalism into ChatGPT for 900M+ weekly active users with attribution and source linking. MUFG (Mitsubishi UFJ Financial Group) also began a formal AI-native transformation engagement with OpenAI in 2026, building on an October 2024 pilot. (OpenAI)

Codex CLI v0.135.0

The May 28 CLI release is the most feature-dense in recent memory:

  • codex doctor now reports environment, Git, terminal, app-server, and thread inventory diagnostics β€” useful for debugging why a session behaves differently across machines
  • /status shows remote connection details and server version when on a remote transport
  • Vim mode gained text-object editing and a configurable interrupt-turn key
  • /permissions understands named permission profiles
  • Non-interactive install via CODEX_NON_INTERACTIVE=1 for CI/automated environments
  • Memory runtime state moved to SQLite (replacing the prior in-memory model)
  • Bug fixes: macOS/Zellij TUI corruption, slash-command completion, resume flows

v0.134.0 (May 26) added search across local conversation history and stronger MCP setup with OAuth support and per-server environment variables. (GitHub Releases)

One active bug to track: Windows OAuth login is broken for some users (reported May 25). auth.openai.com is reachable but the device auth flow fails. No fix merged yet. (OpenAI Community)

Model Updates and Compliance

GPT-5.5 Instant (May 28) got a style update: responses are more conversational and less bullet-heavy. One notable removal: canvas is no longer available in GPT-5.5 Instant or GPT-5.5 Thinking β€” writing and coding features now surface through inline blocks in chat. The Canvas removal was not formally explained in the release notes; legacy GPT-4.5 users retain Canvas access until that model’s June 27 retirement date. In the API, GPT-5.5 is priced at $5.00/M input tokens and $30.00/M output tokens, with a 1M-token context window.

Codex app on Windows now supports computer use β€” the same capabilities previously limited to Mac (Appshots-style screen reading and remote control) are now available on Windows 11, alongside ChatGPT mobile integration. (Thurrott)

Model retirements (ChatGPT only, API unaffected): GPT-4.5 retires June 27; OpenAI o3 retires August 26.

FedRAMP Moderate authorization is live for ChatGPT Enterprise and the API, opening the door for U.S. federal agency use. GPT models, Codex, and Managed Agents are now also available in AWS environments, simplifying deployment for enterprises standardized on AWS infrastructure.

Secure MCP Tunnel (May 19 GA) allows Codex and ChatGPT web to reach private or on-prem MCP servers through a customer-hosted tunnel with no inbound firewall changes required. Initial availability is account-led.

By the Numbers

Codex passed 4 million weekly active users this month. Usage in India surged 27-fold since January 2026; global daily interactions rose 20-fold by late April. The MCP ecosystem hit 9,652 published server records as of May 24 β€” a number that reflects how quickly the tooling layer around OpenAI’s agents is maturing. (New Indian Express, Digital Applied)

What to Watch

  • Appshots for Enterprise β€” Currently Edu-only. Enterprise GA will bring locked-screen remote desktop capabilities to enterprise Codex workspaces.
  • Model retirements: GPT-4.5 sunsets June 27; OpenAI o3 sunsets August 26. Both are ChatGPT-only β€” API is unaffected, but check if you’re using either model in consumer-facing workflows.
  • Realtime API + MCP bug β€” gpt-realtime-2 drops the post-tool response after an MCP call. Community thread is open; watch CLI and API changelogs for a fix.
  • Assistants API migration β€” No new retirement date announced this week, but the direction is clear. If you’re building on Assistants API today, the Responses API is where OpenAI is investing.
  • Windows CLI OAuth fix β€” Outstanding since May 25. Monitor the openai/codex GitHub releases.
  • DeployCo competitive implications β€” McKinsey and Capgemini are founding partners in OpenAI’s new enterprise deployment JV. If your organization uses either firm for AI strategy work, clarify their role as advisors vs. resellers vs. co-deployers.
  • EU AI Act compliance deadlines β€” OpenAI’s Frontier Governance Framework publication suggests the August 2026 GPAI obligations are being tracked. If your organization uses OpenAI APIs in EU-regulated contexts, review whether your usage patterns fall under Annex III high-risk classifications.

That’s the Codex Weekly for May 29, 2026. A foundational SDK redesign, GA milestone for Workspace Agents, and a steady expansion of the enterprise partner ecosystem β€” this week confirmed that Codex is moving from developer tool to enterprise platform.

If your team is evaluating the new Agents SDK architecture, rolling out Workspace Agents, or navigating a Codex Enterprise deployment, reach out to Big Hat Group β€” we’ll help you move from architecture decisions to production. We specialize in AI automation and agent deployment for enterprise environments.

Missed last week? Catch up on Codex Weekly May 15.

Check back next week.