Two Years of Building AI Agents: Producing Answers vs. Producing Right Answers
Tiwaryswarnim · reddit · 2026-09-23
A practitioner shares hard-won lessons from years of building AI agents:
- Start from the end goal, not the model: define the outcome, required information, boundaries, and when the agent should say "I don't know."
- More context isn't better: agentic work is really a context problem in disguise — filter and structure only what the task needs.
- Don't use an LLM as a hammer: prefer deterministic approaches; reserve LLMs for genuine reasoning, interpretation, and flexibility.
- It's art as much as science: how much freedom and context to give is learned through trial and error.
Key point: unlike traditional code, LLMs can fail with a perfectly convincing wrong answer — reliable refusal is a different problem from producing output.
More from coding & agent
- Anthropic's Opus 5.5 system card: multi-agent runs match single-agent outcomes faster — jyangballin · 2026-09-23
- Stripe demos WebMCP integration: a deep dive through a checkout page's new tools — jeff_weinstein · 2026-09-23
- Building AI agent harnesses that improve across 6 editable control dimensions — MaryamMiradi · 2026-09-23
- Drop launches: rootless Linux sandbox that isolates coding agents at the OS level — aronchick · 2026-09-23
- Splitting Reasoning From Tool Use: Two Lightweight AI Models Screen 30,000 Crystal Candidates in 48 Hours — bravo_abad · 2026-09-23
- Sentry Bets Big on Evals: 1,800 PR-Triggered Runs, 20-50% Token Cost Cuts — zeeg · 2026-09-23