CHAPTERS
- 0:00 – 1:36
From autocomplete to agentic work: how engineering workflows changed fast
The speakers look back a few years to describe how AI support in engineering evolved from simple typeahead/autocomplete to agents capable of producing larger chunks of work. They frame the shift as moving from a human-in-the-loop typing line-by-line to coordinating many AI “workers” in parallel.
- •Early AI in coding felt like typeahead/autocomplete assistance
- •Progression from single lines to whole functions, files, and features
- •Two rapid “leaps” in capability over roughly two years
- •Shift from one person doing everything to one person coordinating many AI contributions
- 1:36 – 2:34
What Claude Tag is: proactive AI that joins the channel and drives work
Claude Tag is introduced as a more proactive, channel-native way to use Claude. Instead of opening a chat and prompting repeatedly, you add Claude to a channel and it knows when to jump in, perform tasks, and follow up over time.
- •Claude Tag works inside channels rather than requiring a separate session
- •Proactive participation: jumps in when needed
- •Can run tasks that take days or weeks and follow up automatically
- •Work is visible to the whole group (“multiplayer”)
- 2:34 – 3:15
Why it works now: long-running autonomy and self-scheduling
The research advances enabling Tag are described as increased autonomy and longer effective working time. Claude can self-schedule follow-ups, turning what used to be a single long task into ongoing work across days, weeks, or months.
- •Models are designed to be more long-running and autonomous
- •Ability to operate for extended periods (e.g., ~16 hours per run)
- •Self-scheduling enables multi-day/week workflows
- •Reliability improved: tasks previously untrusted now succeed much more often
- 3:15 – 3:32
Real usage: long-running experiments, daily check-ins, PRs and bug fixes
A concrete picture emerges of Tag sessions persisting for long periods, monitoring experiments, checking data, and sending pull requests when issues appear. The workflow becomes more like ongoing project maintenance than one-off Q&A.
- •Tag sessions can persist for days, weeks, or longer
- •Daily monitoring and reporting on experiment/data health
- •Automatic bug fixes and PR creation as issues are detected
- •“Pull requests just show up” alongside regular data readouts
- 3:32 – 4:25
Memory as the second breakthrough: retaining instructions and channel preferences
Memory is highlighted as a major product and research unlock that makes Tag feel usable and trustworthy. Claude can remember channel-specific instructions and evolve its scope as teammates adjust guidance over time.
- •Memory was difficult to “crack,” but now feels reliable
- •Remembers instructions over time, not just within one prompt
- •Channel-specific preferences (e.g., monitor only certain issue types)
- •Scope can be updated by others and persists going forward
- 4:25 – 4:53
Staying helpful, not annoying: EQ, discretion, and adjustable involvement
They address a common concern: a proactive bot could become noisy. Claude is trained to judge when to participate, can take a back seat, and can be coached to intervene more or less—then remember that behavior.
- •Concern: Tag might jump into every thread unnecessarily
- •Claude trained for situational awareness and restraint
- •Users can correct behavior (more/less proactive)
- •Behavior preferences are remembered for future interactions
- 4:53 – 5:33
Customer patterns and emergent workflows: channel automation that sticks
They share patterns observed in internal and customer use: writing PRs, debugging production, and doing data analysis by connecting Tag to tools. Teams also invent lightweight “rules” for channels (e.g., auto-checking answered questions) that Tag reliably enforces.
- •Common uses: PR drafting, production debugging, data analysis
- •Tool integrations are essential (connect Tag to your systems)
- •Emergent behaviors: enforcing channel norms (e.g., check off answered Qs)
- •Dedicated channels where Tag becomes the first responder or fixer
- 5:33 – 5:47
From “tag it” to “always respond”: hands-off assistance via remembered defaults
The speakers describe a transition from manually tagging Claude on each question to instructing it to respond automatically. This highlights how memory changes interaction cost—turning repeated prompting into durable workflow configuration.
- •Initial workflow: explicitly tag Claude for a first pass
- •Over time: tell Tag to respond every time without being summoned
- •Memory turns repeated actions into stable defaults
- •Result: less overhead and faster help in routine channels
- 5:47 – 6:42
Best-practice diffusion: learning AI in public, shared channels
Because Tag works in public channels, people can observe how power users interact with it and replicate those patterns. This creates rapid internal diffusion of effective prompting, scoping, and workflow setups across teams.
- •Public channels make usage observable to everyone
- •Non-experts learn by watching expert interactions
- •Patterns spread across projects and teams quickly
- •Novel dynamic: shared, social learning for AI tooling
- 6:42 – 7:58
Why it changes team workflows: objective-based, multiplayer AI embedded in collaboration
They contrast Tag with reactive tools that require opening an app and then copying outputs to teammates. Tag takes higher-level objectives, executes continuously, and invites multiple stakeholders to guide the work toward a better final outcome.
- •Reactive tools require remembering to open and prompt
- •Tag accepts higher-level objectives and executes continuously
- •Multiplayer guidance: multiple people steer the same workstream
- •Team nudges improve PR quality beyond one person’s perspective
- 7:58 – 8:14
Lowering the barrier to contribution: non-engineers can ship changes through Tag
Tag is positioned as a way to let more people contribute directly without needing to master terminals, git, or local setups. The conversational interface becomes a bridge to making real changes in the codebase.
- •Goal: broaden codebase contributions beyond traditional engineers
- •Removes friction of terminals, git workflows, and checkouts
- •Conversation-driven requests can yield real PRs
- •Makes collaboration more inclusive across roles
- 8:14 – 8:58
Operational requirements and self-serve knowledge: public channels + source-of-truth access
They emphasize the importance of public channels for monitoring projects and enabling status reporting. Tag also supports self-serve onboarding and policy/HR/legal questions when connected to internal documentation and sources of truth.
- •Public channels enable monitoring across many projects
- •PM workflow: daily status reports across 5–10 features
- •Self-serve answers for onboarding and common policy questions
- •Requires connection to internal source-of-truth files/tools
- 8:58 – 10:48
Productivity impact and choosing Tag vs Claude Code: asynchronous delegation at scale
They share internal productivity metrics—Tag authors a large and growing share of PRs—and explain why: you delegate and move on while it works. They then discuss how a power user decides when to use Tag versus Claude Code as Tag expands from simple fixes to complex, verifiable tasks.
- •Internal impact: Tag responsible for a majority share of PRs (example: ~65%)
- •Key benefit: delegate, then continue other work while Tag runs
- •Initial use cases: small UI fixes, simple data questions
- •Expanded use: complex tasks with verification in a sandbox and tailored per-channel checks
- 10:48 – 11:25
What’s next: more collaboration platforms and org-level customization
They close by outlining expansion beyond Slack (e.g., Microsoft Teams) and emphasizing Tag’s role in transforming team and org workflows. Customizability is a core theme: organizations will adapt Tag to their unique processes and norms.
- •Expansion beyond Slack to other collaboration platforms
- •Goal: keep Claude within “arm’s length” for knowledge workers
- •Focus on team/org transformation, not just individual productivity
- •High customizability to fit different organizational workflows
