Flowing with agents (Interview)

What It Covers

Bian Liu from Sourcegraph discusses AMP, an agentic coding tool that uses multiple LLMs in a for-loop architecture. The conversation explores agent workflows, context window management, token efficiency, and the emerging skill of effective agent interaction through document-driven development patterns.

Key Questions Answered

• •Agent Architecture Fundamentals: AMP operates as a for-loop wrapping agentic LLMs where user input feeds the model, receives tool calls and responses, executes those tools, feeds results back iteratively until completion. This four-step loop architecture forms the foundation of every coding agent, with differentiation coming from tool selection, prompts, and domain-specific sub-agents.
• •Context Window Management: Thread length directly impacts quality, latency, and cost. Quality degradation begins around 70k tokens with severe drops past 120k tokens. Users should treat threads like functions—short, targeted tasks rather than 200-message marathons. Starting fresh threads for each discrete task maintains clean context and prevents model confusion from accumulated irrelevant information.
• •Document-Driven Development Pattern: The Project Enhancement Proposal workflow structures agent interaction through numbered PEPs stored in an admin folder. Each PEP contains status, completion reports, and knowledge base articles. This approach enables asynchronous agent babysitting—checking progress every 10-15 minutes while handling other tasks, rather than constant screen monitoring for optimal productivity.
• •Token Cost Optimization: Senior engineers create short, targeted threads while novice users generate 200-message threads filling context windows unnecessarily. Usage-based pricing reflects actual model costs without artificial rate subsidies. Weekend side projects typically cost under one hundred dollars monthly, while heavy daily usage reaches low hundreds—comparable to dining out expenses for significant productivity gains.
• •Model Quality Variations: Anthropic recently deployed quantized Claude versions causing confirmed quality degradation. AMP mitigates this through multiple inference providers, allowing instant switching when one provider shows degradation or downtime. The system uses different model families for specific capabilities rather than exposing model selection to users, treating it as implementation detail rather than user choice.