OpenAI''s "New Ways of Working with Codex" workshop opened day one of AI Engineer Code Summit in San Francisco on November 10–11, 2026. Coverage from the floor — including a same-day walkthrough of demos by Dominik Kundel and Charlie Guo — centers on three habits: steer with voice, hand work to Codex Cloud, and start from dots when context already lives there.
The product substrate behind those habits is already documented: Cloud environments on ChatGPT Learn and Using Codex Cloud in the Help Center. Together they describe publish-once setups, isolated per-task VMs, and cross-device follow-up without re-running install steps.
TL;DR
| Topic | What to know |
|---|---|
| Cloud friction | Reusable environments cut repeat setup; tasks stay isolated |
| Local ↔ cloud | Agents can hand off branches and spin sibling cloud threads |
| Voice | Dictation plus control-plane steering across threads |
| dots | Context-rich task start, Slack delegation, commute-time PR prep |
| Official docs | Cloud environments guide |
| Limits | See Pro 200 and 28-day Codex push for meters |
Why the summit matters for harness users
Codex, Claude Code, and Cursor-style agents converged on cloud executors in 2026, but mental models lag. Developers still ask whether cloud tasks share one dirty working tree, whether local uncommitted state survives a handoff, and whether voice is a gimmick or a real coordination layer.
The Code Summit session is not a new model launch. It is workflow doctrine from OpenAI''s Developer Experience team: how they want power users to combine Codex Cloud, desktop Codex, dots, and ChatGPT surfaces after the DevDay-era Cloud refresh we tracked in Enterprise Marketplace coverage.
If you run agent benchmarks, note that WorkflowArena already scores orchestration quality separately from worker model strength. Summit advice pushes humans toward orchestrator behavior too.
Codex Cloud: published environments and handoffs
OpenAI''s Cloud environments documentation states the core loop:
- Create an environment with repos, tools, dependencies, and access.
- Let Codex inspect, install, and test the workflow.
- Publish the prepared filesystem.
- Start new tasks from that snapshot in separate workspaces.
Key lines for builders:
- New task in a published environment starts from the prepared filesystem.
- Existing task keeps its own saved files, including uncommitted work.
- Republishing changes future tasks, not retroactively rewriting active ones.
That answers a common failure mode: teams treat cloud like one shared dev container and lose track of which task owns which patch.
Summit quotes add the local bridge: engineers describe asking local Codex to push a branch and continue in a cloud thread, or to spawn cloud work while desktop Codex stays the supervisor. Architecturally, cloud becomes "another executor" rather than a separate product — aligned with how Codex Auto-Review going free reduced quota anxiety on the harness side.
Microphone and chat bubble illustrating voice-directed Codex agent control at AIE Code Summit
Appshot and Sites context
Workshop demos reportedly used a fictional Blossom Music site with live edits. OpenAI also documents ChatGPT Sites for lightweight deployable UIs built through Codex. Appshot-style saves (preserving an app experience inside ChatGPT) showed up as continuity proof when reopening cloud tasks.
For frontend teams, that connects to agentic frontend workshops OpenAI ran earlier in 2026 — Sites are not a replacement for production CI, but they shrink the loop from idea → deployed URL inside the same workspace where Codex runs.
Voice as dictation and as control plane
Summit coverage attributes two voice modes to OpenAI engineers:
Fast dictation while iterating UI copy or layout — low friction for ambiguous creative specs.
Control plane use — checking multiple threads, opening new tasks, pulling context from other apps, and delegating without typing context switches.
One quoted pattern: talk to dots during a commute so PRs await review at the office, because dots already sees dev boxes, laptop state, and cloud environments.
That only works if secrets and repo access in dots match your security model. Compare with Claude Code privacy mods and ChatGPT desktop and Codex tab changes when you wire voice-first workflows.
One-click cloud deploy icon for Codex Cloud tasks that keep running after you close the laptop
dots: awkward for power users, useful anyway
Engineers on stage admitted dots feel awkward next to keyboard-first Codex. The summit still listed three durable roles:
- Safety net — surface misses before merge.
- Delegation sink — especially Slack-driven chores.
- Ubiquity — meet users on phone, desktop, and cloud.
A quoted habit: start from dots, not the Codex prompt box, because dots already encodes calendar, repos, and standing tasks.
That sits next to dots email-safety reporting and Pro 200 meter cuts: delegation is powerful when permissions and usage pools are understood.
Published environments: sharing and enterprise controls
OpenAI''s Cloud environments guide adds team rules worth copying into your runbook:
Privacy → Who can use can scope an environment to a workspace so colleagues start from identical repos, dependency pins, and secret scaffolds. That is closer to golden devcontainer practice than to one shared SSH box.
Only me keeps an environment private while you iterate. Publish freezes a filesystem snapshot; Republish updates the template for future tasks only.
Enterprise admins can gate whether ChatGPT Sites and cloud tasks are available at all — see Enterprise Marketplace listing for how procurement teams think about Work + Codex + dots bundles.
If you already run American Express–style agent rollouts, published environments are how you stop every engineer from reinstalling different CUDA stacks before agents touch production repos.
Local-to-cloud mental model (step by step)
Summit demos emphasized this sequence:
- Start a local task while latency-sensitive (UI iteration, secrets on laptop).
- When the branch is ready for long tests, ask Codex to push and open a cloud thread on the same logical feature.
- Follow progress from phone or browser while the laptop sleeps — the Help Center describes cloud tasks as continuing independently of local machine state.
- Review diffs in the cloud UI before merge; treat cloud like CI with an agent author, not like shared vim.
Failure mode: developers assume uncommitted local buffers magically appear in cloud. OpenAI''s docs say existing tasks retain their own files — you must explicitly hand off via branch push or equivalent agent instruction.
Voice workflows that are not gimmicks
Beyond dictation, summit quotes describe thread management:
- Ask status on task A while task B compiles.
- Spawn a new cloud task without returning to keyboard.
- Brief dots with spoken context before entering the office.
Security checklist for voice:
- Disable voice on repos with production credentials in environment templates.
- Use separate ChatGPT workspaces for client work vs personal experiments.
- Log dots-initiated PRs in the same SOC2 trail as human commits — see Notion transcript exposure for what happens when agent-side logs leak.
How this compares to Claude Code and Cursor cloud paths
| Surface | OpenAI Codex (Nov 2026) | Typical competitor pattern |
|---|---|---|
| Cloud unit | Published environment + isolated task VM | Often single remote workspace |
| Personal agent | dots across Slack/phone | Claude Duet / IDE sidebars |
| Voice | Control plane + dictation | Less emphasized publicly |
| Handoff story | Local agent spins cloud sibling | Varies by harness |
None of these rows pick a winner — they tell you where migration pain will show up when you switch harnesses mid-project. Claude Fable 5.2 ownership guide and WorkflowArena orchestration scores matter more for quality than voice demos.
Practical setup checklist
Use this before you copy summit demos into production:
| Step | Action |
|---|---|
| 1 | Publish a Cloud environment per repo family, not per engineer |
| 2 | Document who can edit vs who can only run tasks |
| 3 | Define when to hand off local → cloud (branch pushed, secrets scoped) |
| 4 | Test voice only on non-production credentials |
| 5 | Track Codex usage separately for dots-initiated runs |
For harness comparison, read Claude Fable vs Opus routing when voice triggers expensive cloud threads.
Metering and limits to watch
The summit did not announce new Pro 200 numbers. Cloud tasks still draw from the same usage pools as desktop Codex unless your contract says otherwise.
Watch these explainx.ai threads when you scale voice + cloud:
- Pro 200 halved October 30 — dots can trigger Codex spend.
- Output token inflation — long cloud tasks amplify hidden tokens.
- Chain-of-thought monitorability — reasoning traces may grow on multi-hour cloud jobs.
Set per-engineer daily cloud caps in procurement reviews, not only seat counts.
What finance and platform teams should ask OpenAI
Before you copy summit workflows enterprise-wide:
- Does dots access SOC2-scoped repos under the same DPA as Codex?
- Are voice transcripts retained, and where?
- When an environment is shared, who owns secret rotation?
- Do cloud VMs egress to arbitrary domains, or allowlists only?
- How do Appshot/Sites deployments inherit SSO from the parent workspace?
Engineer tips quoted at the workshop
Closing advice in summit writeups included "turn it down" (lower reasoning effort when appropriate) and "ask your agents for help" when stuck. The meta-tip: do not cement one workflow — Codex shipped heavily between June World's Fair and Code Summit 2026, per engineer comments that the feature list did not fit one slide.
Related reading
- Codex 28-day daily improvement pledge
- ChatGPT Enterprise on AWS Marketplace
- WorkflowArena orchestration benchmark
- Point72 Codex trade-secrets lawsuit context
- eBay blocks third-party shopping agents
- Loop engineering with coding agents
Cloud documentation URLs and summit timing reflect sources checked November 11, 2026.