Workspace Agents and Codex Slack Integration: From CLI Automations to Team-Shared Agentic Workflows

On 22 April 2026, OpenAI launched workspace agents — shared, persistent agents powered by Codex that run in the cloud and integrate directly into Slack, Salesforce, and the ChatGPT sidebar¹. Two days later, Codex crossed four million weekly active developers². For teams already running codex exec pipelines and CLI-based automations, these announcements raise an immediate question: how do workspace agents relate to the CLI workflows you already have, and when should you use which?

This article maps the architectural relationship between Codex CLI, Codex Cloud, and workspace agents, then provides a practical decision framework and migration guide.

What Workspace Agents Actually Are

Workspace agents are an evolution of Custom GPTs, rebuilt on the Codex harness rather than the Assistants API¹. Each agent gets its own persistent workspace with access to files, code, connected apps, and cross-session memory³. Unlike a Custom GPT, a workspace agent can:

Write and execute code in a sandboxed cloud environment
Use connected third-party apps (Slack, Jira, GitHub, Salesforce, Google Drive) via the plugin ecosystem⁴
Schedule future work and wake up autonomously to continue multi-day tasks⁴
Retain information across sessions, building organisational knowledge over time¹

Workspace agents are available in research preview on ChatGPT Business, Enterprise, Edu, and Teachers plans¹. They are free until 6 May 2026, after which credit-based pricing applies¹.

The Architecture: Three Layers of Codex

To understand where workspace agents fit, consider the Codex platform as three concentric layers sharing a single App Server protocol⁵:

graph TB
    subgraph "Layer 1: Local"
        CLI["Codex CLI<br/>codex / codex exec"]
        IDE["IDE Extensions<br/>VS Code / JetBrains"]
    end
    subgraph "Layer 2: Cloud"
        APP["Codex App<br/>Desktop / Web"]
        CLOUD["Cloud Tasks<br/>Remote sandboxes"]
    end
    subgraph "Layer 3: Team"
        WA["Workspace Agents<br/>Shared, persistent"]
        SLACK["Slack Integration<br/>@Codex mentions"]
    end

    CLI -->|"App Server<br/>JSON-RPC"| CLOUD
    IDE -->|"App Server<br/>JSON-RPC"| CLOUD
    APP --> CLOUD
    WA --> CLOUD
    SLACK -->|"Creates tasks"| CLOUD

All three layers ultimately execute against the same cloud sandbox infrastructure and share the same model roster (gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex)⁶. The differentiation is in who triggers the work and how results are shared.

Dimension	Codex CLI	Cloud Tasks	Workspace Agents
Trigger	Developer in terminal	Developer in app/web	Anyone via ChatGPT or Slack
Execution	Local sandbox (bubblewrap/seatbelt) or cloud	Cloud sandbox	Cloud sandbox
Visibility	Single developer	Single developer	Entire team/workspace
Persistence	Session transcripts, local memory	Cloud task history	Cross-session memory + scheduled wake-ups
Configuration	`config.toml`, `AGENTS.md`, hooks	App settings	Natural-language agent builder
Best for	Tight dev loops, CI/CD, scripting	Long-running tasks, PR creation	Shared team automations, cross-functional workflows

The Slack Integration: @Codex in Your Channels

The Slack integration is the primary surface through which non-CLI teams interact with Codex’s cloud infrastructure⁷. Setup requires:

A ChatGPT Plus, Pro, Business, Enterprise, or Edu subscription⁷
A connected GitHub account with at least one configured environment⁷
Installation of the Codex Slack app from workspace settings (may require admin approval)⁷

Once installed, any team member can mention @Codex in a channel or thread with a prompt⁷. Codex:

Reacts with 👀 to acknowledge receipt
Reviews accessible environments and selects the best match⁷
Creates a cloud task, running against the default branch of the matched repository⁷
Posts results (and optionally an inline answer) back to the thread⁷

Users can specify a repository explicitly — @Codex fix the flaky test in openai/codex — or let Codex infer from context⁷. Enterprise administrators can disable inline answer posting, limiting responses to task links only⁷.

Environment Selection Mechanics

The automatic environment selection follows a simple heuristic: Codex scans the user’s accessible environments, matches based on repository names mentioned in the prompt or thread, and falls back to the most recently used environment if ambiguous⁷. For teams managing multiple repositories, maintaining an accurate repository map in each environment is essential.

When CLI Developers Should Care

If you have been building automations with codex exec, you have effectively been creating single-developer workspace agents. The patterns translate:

Pattern 1: CI/CD Gate → Shared Review Agent

CLI approach:

codex exec "Review this diff for security issues" \
  --output-schema ./review-schema.json \
  -o ./findings.json

Workspace agent equivalent: Create a workspace agent named “Security Reviewer” with instructions to review PRs against your security checklist. Team members invoke it from Slack: @Codex review the latest PR on backend-api.

Pattern 2: Scheduled codex exec → Automations

CLI approach (cron):

# crontab entry
0 8 * * 1 codex exec "Summarise last week's commits and open PRs" \
  --model gpt-5.4-mini -o /tmp/weekly-digest.md

Workspace agent equivalent: Build an agent with automation scheduling built in. Codex can now schedule future work for itself and wake up automatically⁴, eliminating the need for external cron infrastructure.

Pattern 3: AGENTS.md + Hooks → Agent Instructions

Your AGENTS.md files and hooks.json configurations encode organisational knowledge. When building a workspace agent, translate these into the agent’s natural-language instructions:

# AGENTS.md (CLI)
## Build
Run `pnpm install && pnpm build`

## Test
Run `pnpm test -- --coverage`
Never commit with coverage below 80%.

## Review guidelines
- Flag any use of eval() or Function()
- Require error handling on all async operations

Becomes the workspace agent’s instructions:

When working on this repository, always run pnpm install && pnpm build first. Run tests with pnpm test -- --coverage and never commit if coverage drops below 80%. During reviews, flag eval() usage and missing error handling on async operations.

Pattern 4: MCP Servers → Connected Apps

If you run MCP servers locally to give Codex CLI access to Jira, Slack, or databases, workspace agents replace that plumbing with first-party plugin integrations⁴. The 90+ plugins launched in April 2026 include Atlassian Rovo (Jira), CircleCI, GitLab Issues, Microsoft Suite, and Neon by Databricks⁴.

The Decision Framework

flowchart TD
    A[New automation task] --> B{Who needs the result?}
    B -->|Just me| C{Latency-sensitive?}
    B -->|My team| D{Technical or cross-functional?}

    C -->|Yes - tight loop| E["Codex CLI<br/>Interactive or codex exec"]
    C -->|No - can run async| F{Need local files?}
    F -->|Yes| E
    F -->|No| G["Cloud Task<br/>Codex App or API"]

    D -->|Technical - devs only| H{Triggered from Slack?}
    D -->|Cross-functional| I["Workspace Agent<br/>Natural language setup"]

    H -->|Yes| J["Slack @Codex<br/>Cloud task from channel"]
    H -->|No| K{Recurring?}
    K -->|Yes| I
    K -->|No| G

Use Codex CLI when:

You need local sandbox execution with bubblewrap/seatbelt isolation
The workflow requires local filesystem access beyond Git repositories
You are iterating rapidly and need sub-second feedback
You need fine-grained control via config.toml profiles, hooks, and approval modes
You are running in CI/CD pipelines where codex exec is already wired up

Use Workspace Agents when:

The automation serves a team, not just one developer
Non-technical stakeholders need to trigger or consume results
The workflow benefits from persistent cross-session memory
You want natural-language configuration rather than TOML and JSON
The task integrates with Slack, Jira, or other connected apps natively

Use Slack @Codex when:

A team member describes a bug or question in a channel and needs a quick code-backed answer
You want Codex to create a PR directly from a conversation thread
Ad-hoc, context-rich requests where the Slack thread is the specification

Pricing Implications for Teams

The credit-based pricing that takes effect on 6 May 2026 adds a new dimension to the CLI-versus-cloud decision⁸:

Model	Input (credits/1M tokens)	Output (credits/1M tokens)
gpt-5.5	125	750
gpt-5.4	62.50	375
gpt-5.4-mini	18.75	113

For Plus/Pro individual users, messages cost approximately 14 credits (gpt-5.5) or 7 credits (gpt-5.4) each⁸. Cloud tasks are available only on Business and Enterprise plans⁸.

CLI users with API keys pay standard per-token API rates instead⁸, which can be significantly cheaper for high-volume codex exec pipelines but without cloud sandbox infrastructure. The trade-off:

CLI + API key: Lowest per-token cost, full local control, no cloud tasks
Business plan: Cloud tasks, workspace agents, Slack integration, credit-based pricing
Enterprise: Priority processing, admin controls, enterprise security⁸

For teams evaluating the switch, monitor your current codex exec token consumption (visible via OpenTelemetry traces if configured) and compare against the credit pricing table above.

Migration Checklist: CLI Automations → Workspace Agents

For teams ready to promote CLI automations to shared workspace agents:

Audit existing automations — List all codex exec scripts, cron jobs, and CI/CD steps that use Codex CLI
Classify by audience — Single-developer automations stay on CLI; team-facing ones are migration candidates
Extract instructions — Convert AGENTS.md, hooks.json, and prompt templates into natural-language agent instructions
Replace MCP servers with plugins — Swap local MCP configurations for first-party plugin integrations where available
Configure environments — Ensure GitHub repositories are mapped to cloud environments with correct default branches
Test in research preview — Workspace agents are still in preview; validate behaviour before decommissioning CLI automations
Set admin controls — Enterprise admins should configure role-based access, answer-posting policies, and credit budgets before the 6 May pricing transition⁷

Limitations and Caveats

⚠️ Workspace agents are in research preview — behaviour may change before general availability¹.

⚠️ No direct CLI-to-workspace-agent migration path exists yet. You cannot export a config.toml + AGENTS.md bundle and import it as a workspace agent. The conversion is currently manual.

⚠️ Slack integration requires GitHub-connected environments⁷. Non-GitHub workflows (GitLab, Bitbucket) are not yet supported for Slack-triggered tasks, though GitLab Issues is available as a plugin for workspace agents⁴.

⚠️ The automatic environment selection in Slack can misfire for teams with many similarly named repositories. Specify the repository explicitly in your @Codex prompt to avoid surprises⁷.

⚠️ Credit pricing details beyond token rates have not been fully disclosed. Per-task overhead costs and scheduled-wake-up pricing remain unconfirmed as of 24 April 2026.

What Comes Next

The trajectory is clear: OpenAI is building towards a unified Codex runtime where the same agent logic runs across CLI, cloud, IDE, Slack, and enterprise platforms via the shared App Server protocol⁵. For CLI developers, this means your investment in AGENTS.md, hooks, skills, and MCP configurations is not stranded — it is groundwork for the team-scale automations that workspace agents enable.

The practical advice: keep using codex exec for developer-facing, latency-sensitive, and CI/CD workflows. Promote proven automations to workspace agents when they need team-wide access. And install the Slack integration today — the free preview window closes on 6 May¹.

Citations

OpenAI, “Introducing workspace agents in ChatGPT,” 22 April 2026. https://openai.com/index/introducing-workspace-agents-in-chatgpt/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷
OpenAI, “Scaling Codex to enterprises worldwide,” 21 April 2026. https://openai.com/index/scaling-codex-to-enterprises-worldwide/ ↩
9to5Mac, “OpenAI updates ChatGPT with Codex-powered ‘workspace agents’ for teams,” 22 April 2026. https://9to5mac.com/2026/04/22/openai-updates-chatgpt-with-codex-powered-workspace-agents-for-teams/ ↩
OpenAI, “Codex for (almost) everything,” 16 April 2026. https://openai.com/index/codex-for-almost-everything/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶
OpenAI, “Codex App Server README,” codex-rs/app-server, GitHub. https://github.com/openai/codex/blob/main/codex-rs/app-server/README.md ↩ ↩²
OpenAI, “Models — Codex,” OpenAI Developers. https://developers.openai.com/codex/models ↩
OpenAI, “Use Codex in Slack,” OpenAI Developers. https://developers.openai.com/codex/integrations/slack ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴
OpenAI, “Pricing — Codex,” OpenAI Developers. https://developers.openai.com/codex/pricing ↩ ↩² ↩³ ↩⁴ ↩⁵