On August 21, 2026, GitHub shipped a public preview of GitHub Copilot inside Slack. Business and Enterprise orgs can now type @GitHub in a channel to ask questions about code, triage issues, investigate failed builds, or kick off an agent session that runs in a secure cloud sandbox and opens a pull request with a link back to the conversation that spawned it. GitHub is also shipping “Slack Code” — a dedicated channel the agent creates per task, so a long-running fix doesn’t clutter the team’s main channel, and teammates can watch diffs and previews as they land.
It’s a genuinely useful feature, well-documented, and shipped cleanly with no rollout drama — credit where due. But it’s also, structurally, a feature Anthropic already shipped, two months earlier, with a materially broader design.
What Copilot-in-Slack actually does#
Per GitHub’s own changelog, the mechanics are straightforward: @GitHub triggers an agent session. That session can answer questions, open a Slack Code channel to work a task, create PRs, and run asynchronously while you’re away — useful for the classic “kick off a fix before a meeting, come back to a diff” workflow. Sessions are visible to the whole team by design, which GitHub frames as a feature: “developers can see how their teammates prompt the agent and learn from each other.” Admins can require additional approval on any PR the Copilot app identity authors before it merges, which is a sane default given the sandbox runs with real repo access.
All of that is on-demand. Nothing in GitHub’s own description of the feature — and nothing in any of the coverage since — describes Copilot deciding, unprompted, that a Slack conversation is relevant to something it knows and choosing to speak up. You ask; it answers. That’s the entire interaction model.
Claude Tag already did the harder version of this#
This blog covered Claude Tag when Anthropic launched it in research preview on June 23, 2026 (article published June 24), and the comparison here is almost too clean. Claude Tag also supports on-demand @Claude mentions — same basic shape as Copilot’s @GitHub. But it also ships an ambient mode: Claude monitors the channels it’s assigned to and decides, on its own, when to intervene — flagging a deadline mentioned in a different thread three weeks ago, following up on a decision nobody executed, connecting a customer complaint in one channel to a fix that already shipped in another. That requires the model to maintain ongoing comprehension of a channel’s history and make a judgment call about relevance, not just respond to a trigger.
Claude Tag also maintains a single persistent identity per deployment rather than a stateless per-invocation session — meaning context genuinely carries across team members and time, not just within one triggered agent run. A morning-shift engineer’s diagnosis and a midnight on-call engineer’s follow-up share the same Claude, with the same accumulated context, without either having to re-explain anything.
Neither of those two capabilities — ambient proactive intervention, persistent shared identity — appears anywhere in Copilot’s Slack launch. What GitHub shipped is Claude Tag’s on-demand mode, roughly two months later, without the part that made Claude Tag the more interesting product.
The pattern this fits#
This isn’t a one-off gap. This blog flagged the same shape of problem when GitHub Copilot’s model picker absorbed Grok 4.5 the same week Nadella promised a “unified app” merging Copilot Chat, GitHub Copilot, Copilot Cowork, and Autopilot into one coherent product “coming this year,” with no firm date. Copilot’s growth pattern all year has been additive rather than architectural: IDE completions, then Chat, then Workspace, then Cowork, then Autopilot, now Slack — each a genuinely useful surface bolted onto the last, none of them unified by a single agentic loop the way Claude Code’s terminal-native model or Claude Tag’s ambient-Slack model are internally coherent from the start.
That additive pattern isn’t inherently bad — meeting developers in Slack is the right instinct, and plenty of teams will get real value from @GitHub triaging issues while they’re in a meeting. But it does mean Copilot is, once again, shipping the shallower version of a capability a competitor already proved out, in the surface a competitor already claimed. Two months is not a long lag in absolute terms. It’s a long lag relative to how fast this specific race is moving, and it’s the second time in three months this blog has documented Copilot arriving at a surface Claude Code’s ecosystem got to first with a more ambitious design — the autopilot-mode-is-still-not-an-agent pattern all over again, just relocated to Slack.
What to actually watch#
Two open questions matter more than who shipped first. First, whether GitHub extends Copilot-in-Slack toward anything ambient — the changelog gives no signal either way, and Slack-native proactive monitoring is a meaningfully harder product to build safely (false positives in a team channel are worse than false positives in a 1:1 chat). Second, whether Claude Tag’s ambient mode survives contact with real enterprise deployments at scale; research preview numbers and general-availability reality don’t always match, and this blog flagged the privacy tension in ambient monitoring explicitly when Claude Tag launched — an AI that decides on its own when to speak up in your team’s Slack is useful infrastructure and a surveillance surface at the same time, regardless of which vendor built it.
For now, the comparison is straightforward: GitHub Copilot got developers a solid on-demand Slack agent in August. Anthropic got them a persistent, proactive Slack teammate in June. Same surface, different ambitions.
Sources:
- GitHub Changelog — The new GitHub Copilot experience in Slack (primary, fetched directly for feature details and availability)
- This blog’s own coverage of Claude Tag’s June 23 launch and GitHub Copilot’s Grok 4.5 unified-app pattern