Quick Overview
This video is a roundtable technical discussion featuring Anthropic technical staff members Thariq Shihipar, Robert Boyce, and Sid Bidasaria. The engineers review their experiences dogfooding Claude Code and Claude Tag over the course of a year, detailing how their internal software development workflows have adapted to rapid model improvements.
Key Points
- 1.Members of the Claude Code technical staff conduct 70 to 80 percent of their daily development work through an agent named Claude Tag inside Slack.
- 2.Rapid progress in AI model capabilities causes software architecture and tool design to shift substantially every few months.
- 3.Engineers have shifted from micro-managing individual tool calls and transcripts to delegating broader goals and reviewing autonomous outputs.
- 4.Automated fan-out workflows use test-time compute to parallelize deep research, bug identification, and adversarial code reviews.
- 5.Claude Code automates the software verification loop by executing tests, capturing UI screenshots, and compiling feedback directly into Pull Requests.
Summary
Members of the technical staff at Anthropic, including Thariq Shihipar, Robert Boyce, and Sid Bidasaria, sit down to discuss how the internal development process of Claude Code has changed over the past year. In the early days, engineers spent substantial effort prompting models line-by-line, approving individual permission requests, and closely tracking every tool call transcript. Today, the majority of their work happens via Claude Tag, an agent integrated directly into Slack that operates with greater context and autonomy.
Robert and Sid describe how their workflow split has shifted, with 70 to 80 percent of their daily tasks happening autonomously through Claude Tag in Slack, while the local terminal interface and desktop applications are used for roughly 20 percent of tasks requiring hands-on micromanagement. Because Claude Tag operates within Slack, it reads product context and team discussions directly, enabling it to make sounder engineering decisions without requiring human engineers to manually supply background documentation.
The team highlights how fast model turnover fundamentally alters product engineering. Sid points out that while traditional software shelf lives last years, AI underlying capabilities shift every two months. An example discussed is the structured to-do list pattern used during the Claude 3.5 Sonnet era to keep models from giving up on long tasks. Within a year, improved model reasoning made explicit to-do lists obsolete, requiring engineers to stay unattached to temporary harness features they build.
The discussion then moves into modern code review and research workflows. Rather than having human engineers spend time nitpicking syntax in Pull Requests, the team relies on fan-out workflows that use test-time compute. A central agent spawns multiple sub-agents to inspect a codebase from different perspectives, perform adversarial validation, and filter out low-confidence warnings before presenting actionable findings to human reviewers.
Finally, the team details how hosted dev loops and verification tooling make autonomous software development feasible. By running Claude Tag in cloud-hosted developer boxes, the agent can write code, run verification tests, generate UI mockups, capture screenshots, and monitor production feedback without needing a local laptop to stay open. Robert and Sid conclude that while low-level tasks like manual CSS tuning or performance engineering are increasingly automated by AI, software engineering has shifted toward higher-level problem formulation and rapid idea prototyping.
Transitioning from Terminal Prompts to Claude Tag
The Claude Code engineering team reflects on how their workflows evolved from micro-managing low-level tool calls and permission prompts to collaborating with Claude Tag, a Slack-native agent. Team members Robert Boyce and Sid Bidasaria estimate that 70 to 80 percent of their development now takes place directly through Slack interactions, while terminal user interfaces and desktop applications are reserved for fine-tuning.
Adapting to Rapid Model Capability Shifts
The engineers explain that AI technology changes on a cycle of approximately two months, making earlier workarounds obsolete very quickly. Features like structured to-do lists, which were once critical for keeping models like Claude 3.5 Sonnet on track across multi-step tasks, have been deprecated as the underlying models developed stronger innate reasoning, memory states, and autonomy.
Fan-Out Architectures and Test-Time Compute
The team discusses how complex tasks such as code reviews, deep research, and bug detection are managed using fan-out workflows. By launching parallel sub-agents to explore possible issues and then running adversarial verification loops, the system spends compute at inference time to filter results into high-confidence summaries for human review.
Verification Loops and Autonomous Dev Environments
To build Claude Tag using Claude Tag itself, the team created hosted remote environments and automated dev loops. The agent independently generates Pull Requests, runs automated test suites, produces UI mockups, and captures verification screenshots to present clean results without requiring developers to monitor raw transcripts.
The Bottom Line
The discussion establishes that modern agentic AI development has evolved from manual prompt steering into high-level orchestration, automated fan-out verification, and autonomous feedback loops. By dogfooding Claude Tag inside Slack, the team demonstrates how rapid capability advancements replace temporary scaffolding with broader autonomous workflows. The conversation leaves open how software engineering roles will continue to adjust as model update cycles compress even further.
FAQ
What is Claude Code and how does the engineering team use it?
Claude Code is an agentic coding tool developed by Anthropic that assists software engineers with coding, bug fixing, automated testing, and code reviews directly within developer workflows and platforms like Slack.
What is Claude Tag and how does it integrate into team communication?
Claude Tag is a Slack-native agent built by the Claude Code team that lives directly within Slack channels, accessing product discussions and team decisions to execute engineering tasks autonomously.
How has the software development lifecycle changed when building with rapidly evolving AI models?
Because AI models improve significantly every few months, temporary scaffolding and harness features, such as explicit to-do lists, quickly become obsolete and get replaced by broader, autonomous capabilities.
How does the team use fan-out workflows and test-time compute for code review?
The team spawns multiple parallel sub-agents to investigate potential bugs across different angles, applies adversarial reviews to verify findings, and aggregates high-confidence insights into a concise summary for human reviewers.
Why did the engineering team move from running agents locally to hosted developer environments?
Hosted cloud environments allow agents to run persistent loops, monitor feedback, and execute builds continuously in the background without being interrupted when an engineer closes their laptop.
Worth watching for
Software engineers, engineering managers, and technical product developers looking to understand how AI labs dogfood autonomous agentic coding tools in production software lifecycles.
- claude-code
- claude-tag
- software-engineering
- ai-agents
- developer-tools
- code-review