Same model, different harnesses: Coding vs. Support vs. Free Agents
hugobowne · x · 2026-08-17
The post argues that the same base model can behave like completely different agents depending on its "harness":
- Coding Agent: Inspects repos, edits files, runs tests, recovers from errors, and works for an hour.
- Support Agent: Limited turns, bounded tools, refund limits, and hand-off rules for sensitive cases.
- Free Agent: With Bash in a sandbox, it writes code to create capabilities beyond its original toolkit.
The core insight is that the model provides underlying capabilities, while the harness determines visibility, actions, duration, memory, and boundaries. Copying a single architecture (like Claude Code) blindly makes little sense; the right harness depends on the job, and scaffolding that was necessary yesterday may become dead weight as models improve.
More from coding & agent
- MIT paper: clinical multi-agent committees cascade shortcuts; only referee agents catch it — MIT · 2026-08-17
- Agent Optimization Hierarchy: Architecture Over Human Review — MikeBirdTech · 2026-08-17
- GitHub Repo: Complete roadmap to become an AI Engineer in 2026 — ZabihullahAtal · 2026-08-17
- Anthropic Releases Dataset of 817 Cybersecurity Skills for AI Agents — mukul975 · 2026-08-17
- MoneyPrinterTurbo passes 105K stars: turn any keyword into HD short videos — harry0703 · 2026-08-17
- Test shows Sol and GLM-5.3 catch critical code flaws; Claude misses them — morgymcg · 2026-08-17