​Coding agents are evolving fast.

​Every few months, entire agent architectures become obsolete. Planning modes disappear. Handoff systems get ripped out. Models absorb more and more of the stack.

​In this session, you’ll build a modern coding agent system, and learn which abstractions still matter in 2026.

​Join Nicolay Gerold (Amp) and Hugo Bowne-Anderson (Vanishing Gradients) for a live, hands-on build.

​Using TypeScript and Pi, we’ll construct a coding-agent harness that can:

• ​Search and navigate large codebases autonomously

• ​Manage context across long-running coding tasks

• ​Execute tools and recover from failures

• ​Coordinate execution loops and agent workflows

• ​Use modern reasoning models without overengineering the harness

​What we’ll cover

• ​Coding-agent architecture: Build a modern coding-agent harness with tools, execution loops, retrieval, and context management

• ​Context engineering: Learn practical approaches to managing context for coding agents, including retrieval, compaction, and instruction persistence

• ​Planning and handoff: Implement planning and handoff systems, understand why they emerged, and explore why newer frontier models increasingly don’t require them

• ​Harness simplification: See which abstractions modern models have “consumed and what infrastructure still matters

• ​Tool execution and recovery: Build agents that can execute tools, recover from failures, and continue working autonomously

• ​The “Kirby” model thesis: Explore how frontier models increasingly absorb orchestration, planning, and reasoning patterns into the model itself

​Who this is for

• ​Engineers building coding agents or developer tools

• ​AI engineers interested in modern context engineering

• ​Builders who want to understand what still matters in agent infrastructure

• ​Developers frustrated by overengineered agent frameworks

​No prior agent experience required.

​Register to join live or get the recording afterwards.