70% to 80% of Claude Code Team's Work is Now Focused on Claude Tag in Slack
The Claude Code team at Anthropic stated in a discussion lasting about 22 minutes that 70% to 80% of their internal work is no longer done on terminals, with daily deliveries primarily relying on Claude Tag in Slack.
The terminal interface and desktop are only opened in two scenarios: refining a specific implementation or needing to micromanage multiple Claude instances. Instead of a task list, the model is given objectives; Claude Tag separates the interface from the internal transcript, with every visible message in Slack being the result of its proactive tool invocation, while internal monologues remain invisible to collaborators.
The shelf life of underlying capabilities has shortened from yearly to about two months. During the Sonnet 3.5 period, the model often abandoned tasks after completing only three out of five, prompting the team to add a to-do list, which improved completion rates; about a year later, this feature was removed as stronger models no longer required this patch. The conclusion is not to get attached to scaffolding one has built.
Many functions in the harness are explicitly described as patches for current model failure modes, which are proactively deleted as the model improves; however, the scale of tasks has also increased, necessitating the creation of entirely different new tools. It took a long time for AskUserQuestion to teach the model to invoke, but the author now more frequently lets Claude directly produce artifacts with mockups for questioning.
The division of code review has been rewritten: human reviewers previously used a few nitpicks to prove "I have read it," which is now handled by Claude; humans only address constraints that the model does not understand, such as why a certain API looks like this or why service boundaries are drawn here. Workflows evolve from code review, first fanning out to find bugs, then performing multi-faceted adversarial validation on each candidate, filtering out false positives and retaining only real issues; the orchestration layer is code, ensuring no items are missed in for loops.
In market mechanisms, this represents a shift in engineering labor from "humans monitoring terminal conversations" to "humans setting goals, with agents executing in collaboration channels" in an event-driven migration. Computing power and seat budgets have shifted from local TUI to cloud sandbox sessions; the beneficiaries are teams that can clearly articulate goals and solidify verification loops through workflows, while the pressured parties are individuals who still tie value to handwritten pixels and line-by-line patches. Funding and resources are directed towards orchestrating multiple agents rather than piling on more terminal plugins.
The discussion also mentioned that Claude Tag has taken on about 65% of the product engineering PRs for the team; performance optimizations and manually replicating multi-layer radial gradients like those in Mac OS 10.4 Aqua buttons are now considered tasks that the individuals involved can no longer perform as well as the model.
Source: Public Information
ABAB AI Insight
The Claude Code team started with a terminal agent developed by Boris Cherny, subsequently reducing system prompts by about 80%, disabling the overly constraining Plan Mode for the model, and structuring skills as folders with scripts and assets rather than pure markdown. Thariq's path transitioned from game startups and presentation design to this group, initially sharing features with Anthropic employees, only externalizing capabilities that passed internal scrutiny. This "self-use, then delete patches, then change forms" cycle has repeated multiple rounds, with the birth and disappearance of the to-do list being two frames of the same logic.
Capital and resource mobilization is not about buying more IDE plugins but about moving conversations into organizational-level sandboxes: Slack threads equal a remote Claude Code session, with repositories, CLAUDE.md, and skills injected during cloning, while personal machines' MCP and local configurations are excluded. After Code Review covered nearly all PRs internally at Anthropic, the proportion of PRs with substantial comments rose from 16% to 54%, with 84% of PRs with over 1000 lines of changes identifying issues, averaging 7.5 problems; orchestration is written in code as workflows, and adversarial validation keeps false positives in front of human eyes. The motivation is that task scales exceed single-window contexts, necessitating the use of isolated sub-agents to maintain clean contexts.
This is analogous to GitHub moving code reviews from email to Pull Requests, and AWS moving machine management from SSH to control panels: the interface is not the product; orchestration power is. The industry position has shifted from "completing the next line" to "controlling the entire development lifecycle"—implementation can still use terminals, while demand, review, on-call, and CIBabysitting have migrated to Tags. Cursor, Cognition, and others use in-editor agents, while Anthropic has split the same model capabilities into terminal implementation layers and channel collaboration layers.
Structural judgment belongs to the technical replacement layered with industry chain reconstruction. It is not the programmer identity that is being replaced, but the low-information review of "proving I have read it" and manual pixel labor; reconstruction occurs as the toolchain shifts from editor plugins to organizational identities plus sandboxes plus workflow runtimes. The mechanism is that model failure modes change every two months, and patch-type harnesses must be deletable, or else outdated constraints will be written into the next generation of models; as tasks grow larger, this compels the emergence of entirely different orchestration layers. Only deleting without building leads to stagnation, while only building without deleting pays protection fees to old failure modes.
ABAB News · Cognitive Laws
Patches have a shelf life; goals yield compound returns.
When interfaces can hide monologues, agents are considered part of the collaboration layer.
Humans no longer prove they have read; they only explain why boundaries are drawn here.