Hugo Bowne-Anderson Warns 1-to-5-Turn AI Agents Need Smaller Harnesses as Models Absorb Features
Updated
Updated · O'Reilly Media · Jul 22
Hugo Bowne-Anderson Warns 1-to-5-Turn AI Agents Need Smaller Harnesses as Models Absorb Features
2 articles · Updated · O'Reilly Media · Jul 22
Summary
Bowne-Anderson argues most AI agents—especially support, sales and enterprise systems resolving tasks in 1 to 5 turns—should use a minimum viable harness instead of coding-agent-style memory, compaction and sub-agent stacks.
The essay frames harness design around two axes: action complexity and context complexity. Builders should add tools, state, routing, guardrails and handoffs only when the job actually demands them.
He defines the “Kirby effect” as frontier models absorbing assumptions baked into harnesses, making features obsolete; Manus was re-architected 5 times in a year, and Anthropic and LangChain also repeatedly rebuilt agent systems.
Coding and deep-research agents still need heavier context management because they run long reasoning loops, but Bowne-Anderson says even those cores can stay small—citing coding and search agents built in 131 and 61 lines of Python.
The broader takeaway is to revisit every harness addition when stronger models arrive, because yesterday’s workaround may become dead weight while fundamentals like evals, tool design and safety remain durable.