Claude Tag — Ambient Loops in Team Tools¶
Claude Tag is Anthropic's product for a persistent, ambient Claude agent inside Slack — distinct from the earlier Claude Code in Slack integration, which still runs on Pro/Max plans under an individual user's account. Rather than a scheduled cron or a manually triggered terminal session, the loop lives inside the communication layer where the team already works.
(Anthropic, "Introducing Claude Tag", Jun 2026.)
Why this Part II doc carries dates, not
v2.1.xversion stamps. Part II promises that platform facts carry the version they were true in. That promise binds facts about the Claude Code CLI, which the changelog versions. Claude Tag is a separate Anthropic product on its own release timeline: the official Claude Code Slack page states no version requirement, and the CLI changelog contains no Claude Tag entry (checked 2026-09-07 against the full file,0.2.21–2.1.263). So the facts below are dated to their announcement instead, which is the strongest provenance available. A trap worth naming: the changelog's/tagentries — "Removed/tagcommand" (v2.1.92) and a/rename-and-/tagsession fix (v2.1.19) — are about naming a session, not about Claude Tag, and must not be cited here.
Architecture¶
Channel-scoped identity¶
Each Slack channel gets a distinct Claude instance with isolated memory and tool access. Sales Claude ≠ Engineering Claude — different permissions, different context, different guardrails. Admins configure each instance per channel.
Ambient context¶
Claude Tag reads the full channel history without requiring re-explanation. It builds tacit knowledge from team conversations over time and can draw on other accessible Slack channels and data sources, though it does not report information out of private channels. This is persistent memory built from natural team communication rather than explicit knowledge files.
Self-scheduling¶
Claude Tag can schedule tasks for itself and pursue projects autonomously over hours or days. It supports ambient behaviour: proactively flagging relevant information and following up on stalled threads without being asked.
Isolated sandbox per invocation¶
Each invoked task spins up its own sandbox — clones repos, writes code, runs tests, commits work. The pattern is identical to the worktree isolation model in Background Agents.
Governance¶
- Token spend limits per org and per channel
- Full audit logging for all agent actions
- Role-based memory separation enforced at the instance level
The Third LLM Paradigm¶
"Imo this is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans." — Andrej Karpathy, X, Jun 2026
| Generation | Form | Interaction model |
|---|---|---|
| 1st | Website / chat | One-shot conversation; user-directed |
| 2nd | App integration (IDE, terminal) | Session-based; user-initiated |
| 3rd | Ambient entity | Persistent; self-scheduling; org-wide context |
Claude Tag is Anthropic's first shipped instance of the third paradigm. Boris Cherny (Claude Code's creator) frames it as "Claude Everywhere": loops embedded across all team tools, not just the terminal.
Signal data¶
- 65% of Anthropic's product team's code is created using their internal version of Claude Tag
- Powered by Claude Opus 4.8
- "We see Claude Tag as the beginning of an evolution of Claude Code: it makes the model even more proactive, and it works better with a full team." — Anthropic, "Introducing Claude Tag"
- A newer, higher figure — but paraphrase, not a direct quote. Lenny's Podcast's own show notes for a Feb 2026 episode with Anthropic CPO Mike Krieger state "90% of Claude's code is now written by AI" and separately tease "90-95% of code for some products" — both are the show's paraphrase (the transcript is paywalled, no verbatim Krieger quote is available), and the two figures describe different, unstated populations — treat as a directional signal above the 65% figure, not a precise update to it. The same notes summarize Krieger's MCP framing as "everything becomes an MCP endpoint, making the entire digital world scriptable and composable by AI agents." (Lenny's Podcast, "Anthropic's CPO: here's what comes next" — Mike Krieger, Feb 2026.)
Production Deployment: On-Call for CI/CD¶
Anthropic runs Claude Tag as a persistent Slack on-call responder for CI/CD incidents, running the same monitor → triage → investigate → verify loop as the architecture above, specialized for infrastructure failures: it spins up parallel executor subagents to investigate each candidate dependency via MCP connectors, then verifies the proposed fix before reporting back.
Speed: Claude posts its first evidence-grounded analysis a median of 14 minutes after an incident opens (measuring time-to-first-analysis, not detection time), and in the fastest observed cases names the root cause within 4 minutes.
Durable memory across incidents (lessons.md): a running log of every incident
resolved — what happened, the root cause, the fix, and the gotcha worth remembering —
that Claude appends to automatically. Every new investigation starts by reading it first.
A notable entry: "query the data first, then theorize. Config tells you what could go
wrong; metrics tell you what did." When the same pattern recurs often enough, it gets
promoted out of the log and into the investigation skill itself — a lighter-weight
version of Warp's inner/outer skill-improvement loop,
here triggered by pattern frequency in a log rather than a scheduled observer agent.
(Anthropic, "How Claude Tag serves as Anthropic's first responder for CI/CD failures", Aug 2026.)
Self-Improving Instructions via a Human-Gated Friday Routine (Legal)¶
Anthropic's legal team runs a recurring Friday routine where Claude Tag reviews the week's counsel feedback and proposes edits to its own shared channel instructions, which a human approves before they take effect — "Claude Tag added this new instruction to its set of instructions to follow in all future reviews, allowing it to improve with channel feedback in real time." A channel-native instance of the same inner/outer loop as Warp's skill split, except the "outer" step edits ambient channel instructions directly rather than a versioned skill file. The same post reports legal review of marketing assets compressed from a day to 30 minutes by having Claude check every asset first against internal documentation and the public web. (Anthropic, "How Anthropic employees use Claude Tag", Aug 2026.)
A Working Practitioner's Own Loop Specs¶
Boris Cherny — Claude Code's creator, whose "I have loops that are running... my job is to write loops" framing is this KB's own starting point (docs/01) — describes Claude Tag writing more than 50% of his own PRs every day, and roughly 100% of his data analysis, running on Opus 5.5 and "Fable 5.1." More useful than the headline percentage are four example prompts, each of which is effectively a loop spec in prose — SCOPE/ACTION/BUDGET/STOP stated informally but all present:
- A reactive trigger: "react to threads in this channel with ✅ when resolved" — watch for a specific condition in a channel and act when it fires.
- An end-to-end repro-then-PR loop: "from now on, attempt to repro every bug in this channel end to end, running the full app. once you reproduce the bug, put up a pr to fix it and tag the right team for code review" — SCOPE and ACTION chained straight through, with the human handoff coming after the PR is up, at code review.
- A budgeted deep-dive: "brainstorm ~100 hypotheses for what could explain this weird data, use a workflow to in/validate. spend 10m tokens or so digging very deep. draw a chart with the result." — an explicit BUDGET in the same breath as the ACTION, rather than left to a platform default.
- An explainer-artifact loop: "make me an interactive game explaining how this part of the code works, then make a slide deck explaining it for others on the team" — a durable explanation artifact rather than a chat reply, closer to a Routine's or a skill's output shape than a one-off answer.
(Prompts: @bcherny, Sep 2026.)
Coming from the practitioner most associated with popularizing "loop engineering," this is evidence the framing is not just descriptive of what other teams build, but of how its own originator works day to day — and at volume. In a Sep 23 reply on a different post (his Lean/Agent SDK thread), asked what he does while agents work: "I usually have at least 10 Tag/Projects sessions running at a time. While the agent works, I start another session or unblock one that is running" (@bcherny, Sep 2026). Reading roughly 85 of the Tag post's ~190 replies (as of 2026-09-28; not the full thread) surfaced two operational details, each its own reply — "Environment runs in the Cloud" (@bcherny, Sep 2026) and "Team has a 2-seat minimum" (@bcherny, Sep 2026) — and one opinion, answering whether they really use it in Slack: "It's surprising, but it feels awesome to have Claude in the same place where we already talk to each other" (@bcherny, Sep 2026).
First-Party Case Study: A Performance Loop Run in Tag¶
Anthropic's claude.ai team ran a two-week performance sprint in August 2026 through Claude Tag (beta), "running an internal research model roughly comparable to Opus 5.5", from one Slack channel with standing instructions that began "Your job is to facilitate all things related to the performance of the claude.ai website and desktop app." The write-up is a worked loop contract:
- Scope. Claude used the Datadog MCP server to find "the four highest-impact user journeys", and each thread was kept "deliberately narrow, focused on one benchmark or journey".
- Verifier. Every benchmark had two jobs: "a metric Claude could move in the lab" and "a guardrail in CI with a number that could only ratchet down". Flaky benchmarks, and ones that did not track user latency, were thrown out.
- Gates. "Every PR went through automated review with at least one human approval". User-visible changes shipped behind short-lived flags ("nearly two hundred flags, more than half of which were already cleaned up by the end"). High-risk changes rolled out to employees first, then to one percent of users, then to everyone.
- Scale. "During the sprint, we ran more than a hundred and fifty at a time" (threads), and more than two hundred changes landed on the busiest days.
- Stop. Humans overrode the model's own instinct to stop: "Let's keep driving this down, the targets are not the stopping point."
Reported result: "more than three thousand changes without a single customer-facing incident or rollback", and "3.1x faster on average (geometric mean)" across thirteen p75 measurements over four journeys. The team's own caveat is the useful counterweight: "The loop was productive, but it wasn't autonomous." Humans supplied ambition, taste and direction. This is Anthropic engineers reporting on Anthropic's own product, not an independent result. (claude.dev, "How we made claude.ai 3x faster in two weeks", Raymond Wang, Sam Attard and Issac G., Sep 2026; announced by @ClaudeDevs and quote-tweeted by @bcherny.)
Deployment mode comparison¶
| Mode | Trigger | Persistence | Context |
|---|---|---|---|
| Terminal session | Manual | Per-session | Local machine |
| Routine | Schedule / API / GitHub event | Cloud-hosted | Configured per routine |
Background agent (--bg) |
Manual flag | Session-isolated | Per worktree |
| Claude Tag | @mention in Slack | Persistent channel memory | Org-wide + channel history |