- Author
- Hugo Bowne-Anderson (Vanishing Gradients)
- Published
- 21 July 2026
- Kind
- Substack essay, with promotion of a paid workshop
- Original
- hugobowne.substack.com/p/stop-overengineering-your-agent-harness
- Subject
- Harnesses and agent loops
- Themes
- Small means
What it is
Hugo Bowne-Anderson writes the Vanishing Gradients newsletter. The essay defines agents and harnesses and then maps harness needs by task. It aims to steer builders away from heavy harnesses they do not need.
What it says
- The essay argues the needed harness depends on action complexity and context complexity.
- The post defines an agent as an LLM with tools in a loop.
- It lists five harness jobs: loop, tool execution, context management, state and safety.
- The post says a coding agent can be built in 131 lines of Python.
- Nicolay Gerold's Kirby effect: harness parts encode model weaknesses and expire as models improve.
- The post says Manus was re-architected five times in a year.
- The post says support agents may need far less memory and compaction than coding agents.
What evidence it gives
The essay offers argument backed by named examples and quotes from practitioners. The Maven Assistant case is one team's reported experience, relayed by the author. It reports no benchmarks or experiments.
Interests and limits
The author runs paid workshops and promotes one in the post, with a discount code. The two-axis model is the author's framing and is not tested. The Kirby effect rests on reports of rebuilds, not measurement. Only two reader comments appear on the page.