Improved Models for Long-Horizon Work

Models have gotten better at staying on track through longer tasks without losing the thread.

Meaningful improvements in tool calling and sustained reasoning across many steps.

The Ralph Wiggum Loop

Models don't need to get smarter to handle long-horizon tasks.

You just need a while loop around them.

How the Loop Works

• Agent is stateless: context resets each cycle
• Progress persists via external memory
• Chips away at tasks until requirements are satisfied
  • AI-as-judge, all tests pass, etc.

Beads: Trello for Agents

A shared notebook your agents can actually read back.

• Tasks and context persisted as JSON "beads" backed by Git
• Dependency tracking, progress tracking, cross-session memory
• No embeddings or vector stores needed

Agent Orchestration: Gas Town

• Break complex work into subtasks
• Coordinate between specialized agents
• Maintain shared context
• Inspired by Kubernetes system design
• Wild west early 2026

OpenClaw: Beyond Coding

A persistent AI assistant that takes real-world actions.

• Long-term memory + multi-agent composition
• Full system integration in a single runtime
• Have your private life leaked

Claude Code: Lean and Pragmatic

• A way to track work items
• The ability to spawn specialized subagents
• Leader and specialist agents coordinate asynchronously

AI coding maturity ladder

Fading Skills

• Syntax memorization
• Boilerplate coding, yet another
  • REST endpoint
  • command handler
  • screen
  • ...

Rising Skills

• Building out walking skeletons and good examples (tm)
• System design and task specification/breakdown
• Validating generated code (the "sniff test")
• Testing and debugging AI workflows
• Agent orchestration, harness engineering, context management and prompting
• Multi-agent coordination