Model vs Harness: Developer Analyzes the Three Schools of AI Agent Architecture
AccBalanced · x · 2026-07-30
Developer @justinsunyt provided a sharp breakdown in response to Martin Casado's musings on the value of agent harnesses, arguing that the three competing beliefs are all correct in different contexts:
- Less is more (Models are the magic): For isolated, single-shot tasks that don't interact with external systems, models benefit more from minimal context bloat. For instance, mini-swe-agent with a single bash tool outperforms Codex and Claude on coding benchmarks while being significantly cheaper.
- Post-training wins: The RL-ification of models leads them to overfit on specific tool schemas (e.g., Opus hallucinating parameters, GPT preferring native applypatch). As labs invent new multi-agent orchestration approaches, spawning subagent behaviors will increasingly depend on post-training.
- Harnesses provide independent value: When you care about agent behavior around interfaces and systems, harnesses offer irreplaceable support. As agents evolve and move towards the cloud, the independent value of harnesses will only grow.
More from AGI Musings
- Gary Marcus Questions Anthropic's $100B+ Revenue Projections Amid Rising Competition — GaryMarcus · 2026-07-30
- Robotics Leaders on How World Models Drive Real-World Deployment — Scobleizer · 2026-07-30
- Elon Musk: Humans Will Lose Control of AI Within Ten Years — rohanpaul_ai · 2026-07-30
- Geoffrey Hinton says AI will replace all mundane intellectual labor; startup raises $75M for AI invoice auditing — EXM7777 · 2026-07-30
- Core of AI is Self-Empowerment: Upskilling Over Outsourcing — HamelHusain · 2026-07-30
- Runtime Scaling Triggers Red Queen Race Across the Economy — QuintinPope5 · 2026-07-30