Stress-testing Instinct agent: snacks, flight alerts work, payments remain a trust blocker
Sad-Locksmith5980 · reddit · 2026-09-08
A first-hand review of running the Instinct agent on real-world tasks beyond summarization:
- What worked: ordering snacks across apps; monitoring flight prices over a date range (3 checks/day) with rise/drop alerts; checking whether parents went for their post-dinner walk (including how to actively ping them); researching startups into concise notes.
- Pain points: noticeably slow workflows with big gaps between steps; limited browser/site access with human-in-the-loop needed for OTPs; and payments as the biggest trust issue — comfortable letting the agent compare and prepare orders, not handing over card details.
- The author opens discussion on which tasks people trust agents with fully autonomously, payment workarounds, and country-by-country differences in access and services.
More from coding & agent
- Developer runs a 'ghost' agent to babysit Codex long-running terminal tasks — BLUECOW009 · 2026-09-08
- The funniest OpenAI snooping theory: two Codex agents found and helped each other — gleech · 2026-09-08
- Tip: disable Astra Ultra heartbeats before bed or they'll drain your Codex quota overnight — jdjohnson · 2026-09-08
- One of the Best Ways to Deploy Agents: Builder, Reviewer and Verifier — BLUECOW009 · 2026-09-08
- Website AI agents work best as lead capture triggered by intent, not on arrival — gogeta7124 · 2026-09-08
- A 2-minute fix burned 100k+ tokens: taming Linux agent memory with MemOS — NutellaEquivalent452 · 2026-09-08