Keep pulling the thread on Thariq Shihipar.
The workflows feature in Claude Code allows users to spin up sub-agents to perform, parallelize, and verify work.
To mitigate self-preferential bias, the recommended pattern is to use separate agents for coordinating, executing, and verifying work.
The Claude Code team reduced the length of its system prompt by 80%.
ClaudeTag has significantly changed how work is done internally at Anthropic.
Continued development and improvement of ClaudeTag is a major focus for Anthropic.
A key benefit of the `workflows` feature is enabling a separate verification agent to review an output against a rubric, which is useful for non-deterministic tasks.
Anthropic has observed a "self-preferential bias" where a language model is more lenient when verifying its own outputs.
Anthropic's ultimate vision for Claude is for it to be a proactive agent that meets users where they work.
Claude Code includes features named /loop, /goal, and workflows, which are designed to enable agents to run for extended periods.
The /goal feature in Claude Code helps an agent remember its exit condition and prevents it from stopping until that condition is met.
The workflows feature in Claude Code was used for the rewrite of the `bond` project in the Rust programming language.
As language models become more capable, they require less explicit direction, fewer constraints, and fewer examples in their system prompts to perform effectively.