OpenAI shipped expanded ChatGPT agent capabilities for workplace tools in April 2026, enabling autonomous AI teammates that can operate natively inside Slack channels and Linear project management workspaces without requiring continuous human prompting. The rollout, delivered through a combination of ChatGPT Apps, the Agents API, and Slack bot integrations, marks one of the most significant expansions of agentic AI into day-to-day team workflows to date. Our read: the three-path framing is more honest than the typical single-product launch — and that specificity is what makes this worth a serious look from any team already running Slack and Linear together.
The capabilities were highlighted as a notable development in a recent video roundup covering AI updates — specifically in a segment titled "China's Free AI Just Embarrassed Claude And ChatGPT (+12 AI Updates)" — which flagged autonomous AI teammates in Slack and Linear as item three among the most consequential near-term shifts for enterprise teams.
Three Product Paths to Autonomous Agents
OpenAI is not marketing these capabilities under a single unified product name. Instead, the company bundles agentic functionality across three distinct channels, each suited to different deployment contexts.
The first path is ChatGPT Apps — custom integrations that live within ChatGPT for Business and Pro tiers. These allow teams to write custom actions, or use OpenAI's action templates, that call Slack or Linear APIs directly. When a user mentions an app in a ChatGPT conversation, it can autonomously invoke those actions. This path is synchronous: the user asks, ChatGPT responds.
The second path is the Agents API, OpenAI's raw agent orchestration layer. Developers provide a system prompt, a set of tools such as the Slack API and Linear API, and an instruction set. The API then runs the agent in a continuous loop — observe state, plan, act, observe again — without requiring a human to prompt each step. This path carries a higher latency, ranging from 30 to 120 seconds per cycle, but is fully autonomous by design.
The third path combines a Slack bot with the Agents API directly. In this configuration, a Slack message triggers a backend function — hosted on a service such as Convex or AWS Lambda — which calls the Agents API. The agent runs in the background and posts its results back to Slack. This architecture supports what OpenAI and integrators are calling 24/7 autonomous operation, within the limits configured by the deploying team.
According to OpenAI's Agents API documentation and Slack's published API and integration pattern guides, teams seeking to trigger agents from within ChatGPT itself should use the ChatGPT Apps path. Teams wanting asynchronous background automation — such as reading a ticket, drafting a pull request description, and posting the result to Slack — should use the Agents API directly or in combination with a Slack bot. Teams requiring always-on monitoring, where a new Linear issue automatically triggers a summary posted to a triage channel, should build on the Slack bot plus Agents API combination.
What "24/7" Actually Means in Practice
The marketing framing of an AI teammate that "works around the clock" has prompted some clarification about what continuous availability means in technical terms. In practice, agent availability is determined entirely by the trigger type configured at deployment, not by any form of continuous background cognition.
Event-driven triggers are the preferred architecture. In this setup, a Slack app subscribes to events such as message_created or reaction_added. The moment a message lands in a designated channel — such as a bugs or triage channel — the agent wakes, reads the surrounding context, and calls the Linear API to create or update a ticket. Latency in this configuration runs between two and eight seconds, and costs are incurred per API call.
Scheduled polling offers a lower-cost alternative. In this model, a Lambda function runs on a fixed interval — every five minutes, for example — and queries Linear for new issues since the last run. If new issues exist, the agent summarizes and posts to Slack. Latency in this configuration can reach up to five minutes plus cold start time, with costs tied to Lambda invocations and ChatGPT API usage per batch of issues.
User-triggered, synchronous operation is the simplest path. A user types a slash command in Slack — for instance, directing the agent to create a design system ticket for a Figma redesign — and the Slack event fires a Lambda that calls the Agents API, with results posted back to the channel within 20 to 60 seconds.
The distinction matters for teams evaluating the technology. In my experience, "24/7" is where most early adopters overbuild — wiring up continuous polling infrastructure when a simple event subscription covers 90% of use cases at a fraction of the cost; the latency and cost profiles above make that tradeoff concrete. "24/7" in this context means the agent wakes whenever its trigger fires, not that it is continuously thinking or monitoring. For true continuous monitoring — watching for issues crossing defined thresholds and escalating automatically — teams require dedicated infrastructure such as a systemd service or Convex cron job, rather than relying on Slack or Linear event subscriptions alone. This clarification comes from Slack's event subscription documentation and OpenAI's published Agents API polling best practices.
Verified Capabilities and Hard Limitations
As of April 2026, verified capabilities of the agent stack include reading Slack channel messages and generating summaries, querying the Linear API to fetch issues by label or description, creating Linear issues from Slack messages with title, description, assignee, and priority fields populated, closing and reopening issues, posting formatted status updates with links and user mentions, generating summaries of Linear epics threaded into Slack, and drafting pull request descriptions based on issue context.
The limitations are equally specific. Agents cannot process audio from Slack Huddles or voice channels — voice transcription requires a separate pipeline built outside the core integration. They cannot read attached files beyond metadata, ruling out document understanding use cases involving Slack file uploads or Linear design assets. Agents also cannot trigger OAuth flows for third-party services such as GitHub, Jira, or Figma; they can only call APIs that have been pre-authorized at deployment time. Real-time participation in Slack Huddles or calls is not supported, with post-meeting summaries being the only available option. Finally, agent memory does not persist across workspaces — each Slack workspace and Linear team pairing is treated as an isolated context with no visibility into other teams or organizations.
These constraints are documented in OpenAI's Agents API limitations guidance and Slack's published app permission model.
Deployment Architecture
The typical deployment stack for an autonomous Slack and Linear agent follows a webhook-driven pattern. A Linear issue creation event sends a webhook to a cloud function hosted on AWS Lambda, Convex, or a similar platform. That function calls the OpenAI Agents API, which runs an observe-plan-act loop using the Linear API and Slack API as available tools. The agent then posts its result — a summary, a status update, or a newly created ticket reference — to the designated Slack channel or as a comment on the Linear issue.
Slack app configuration requires enabling specific permission scopes: chat:write for posting messages, channels:read for reading channel metadata, users:read for accessing user profiles, and optionally reactions:read for message reaction handling. Linear API access is provisioned through a dedicated API key generated in Linear's settings dashboard, which OpenAI and security best practices recommend storing in a secrets management service such as AWS Secrets Manager rather than in environment variables directly.
Developers building on this stack can choose among AWS Lambda for the most common enterprise deployment pattern, Convex for tighter integration with Next.js-based products, or Replit for rapid prototyping with a local webhook tunnel via ngrok.
Competitive Context
The April 2026 release positions OpenAI's agent offerings directly against Claude MCP-base
Written by Hiram Clark, Editor — vybecoding.ai
Published on May 1, 2026