One agent per user in a microVM: lessons from running a proactive AI friend in production
maritime_sh · reddit · 2026-09-03
A developer shares production lessons from a language-learning product where each user gets an AI friend that texts first from a real phone number, built as one agent per user in its own microVM rather than a shared model with a user ID.
- Memory solved by architecture: friend history lives inside the agent—no vector store, no retrieval; idle agents cost nothing, wake in 1 second, and signup-to-first-message is 37 seconds.
- Real machines can act: agents generate selfies, own phone numbers and emails, and call a scoped API to change user settings (e.g., self-setting do-not-disturb); per-agent token scoping limits blast radius to one account.
- Persona is not a security boundary: agents leak runtime, model, OS and workspace details in character-breaking detail; hardening the persona failed, so outputs are filtered and replaced with in-character lines.
- Honest failure: 30 proactive messages daily get almost no replies—getting humans to answer unprompted messages remains unsolved.
More from coding & agent
- JamesDSP breaks stereo on T2 MacBook 6-channel speakers — fixed with Copilot CLI's help — DanWahlin · 2026-09-03
- Why aren't game studios using agent swarms to remaster old classics? — CSProfKGD · 2026-09-03
- New CSS progress() trick simplifies adaptive border radius — jh3yy · 2026-09-03
- Team shares 4 real LLM uses: contract negotiation, agent clarification, grading, math — xuanalogue · 2026-09-03
- RealSWE benchmark: realistic user requests test coding agents, explicit intent boosts results — skku · 2026-09-03
- Grok Build v1.0.18 adds managed MCP policy enforcement, multi-round tool elicitation and charts — XFreeze · 2026-09-03