AI Insights · Oct 22, 2026 · Verified
From "start simple" to "own the boundary"
In 2024 the labs told you to start with raw APIs. In 2026 the advice is to control the edges: keep checking separate from generating, put gates on tools, and strip out scaffolding as models improve.
Claim
The 2024 and 2025 guides are still live and unrevised: start with the API, use the simplest thing, build an agent only for complex decisions, hard-to- maintain rules or unstructured data, single agent before multi-agent. The 2026 lesson layered on top is that the loop itself is a commodity. What you own is the boundary: permissions, typed high-stakes tools, an independent evaluator, durability.
Evidence
- Anthropic (2024-12-19): "start by using LLM APIs directly"; workflows are predefined code paths, agents direct their own process; "find the simplest solution possible". The page carries a note that tooling has changed but the text stands.
- OpenAI guide (2025): three reasons to build an agent; otherwise "a deterministic solution may suffice"; evolve to multi-agent "only when needed"; two human-intervention triggers, failure thresholds and high-risk actions.
- Anthropic (2026-03-24): prompt and harness engineering "eventually hit ceilings"; agents "reliably skew positive when grading their own work"; a separate skeptical evaluator driving the app with Playwright is "far more tractable"; solo Opus 4.5 produced a broken app in 20 minutes for $9, the full harness a working one in 6 hours for $200; sprint contract removed for Opus 4.6.
- Anthropic (2026-04-02): build on tools the model knows (bash plus editor reached 49% SWE-bench Verified with Claude 3.5 Sonnet); remove scaffolding the model no longer needs; promote high-stakes actions to typed declarative tools so the harness can gate, log and replay; ask "what can I stop doing?".
- Claude Code docs: agent teams use about 7x tokens; disabled by default.
Open questions
- Whether the evaluator should be a cheaper model than the generator.
Sources
- https://www.anthropic.com/engineering/building-effective-agents — blog, Wed Dec 18
- https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf — paper, Mon Mar 31
- https://www.anthropic.com/engineering/harness-design-long-running-apps — blog, Mon Mar 23
- https://claude.com/resources/articles/harnessing-claudes-intelligence — blog, Wed Apr 01
- https://code.claude.com/docs/en/costs — docs, Wed Oct 21
AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.