Agents Should Be More Opinionated (www.vtrivedy.com)

A recent position piece argues that the best AI agents aren’t the most flexible—they’re the most opinionated. Instead of exposing users to endless knobs (temperature, chunking, model choice), builders should bake decisions into an agent harness: curated prompts, tool sequencing, context management and subagents that encode what “good” looks like. The result is a product that works reliably out of the box, reduces user friction, and speeds value delivery—backed by rigorous dogfooding and task-focused evals rather than generic benchmarks. Technically this reframes agent design: an agent = opinionated harness + model. Models are “spiky” and non-fungible—behavior depends on the harness—so upgrades can break finely tuned prompts or tool integrations. The practical implications: prioritize harness engineering over generality, start deep-and-narrow (optimize a single workflow), hardcode well-known defaults, and evaluate model+harness pairs on real tasks. Examples cited include industry teams and products that bake domain-specific tooling and defaults into their agents. For builders, the takeaway is clear: accept tradeoffs, make decisions up front, and iterate via dogfooding and evals—opinionated defaults win reliability and user trust, with customization layered on later.