'The Session You Cannot Take With You': Six Concrete Ways Inference APIs Have Quietly Made Agent Transcripts Non-Portable
Earendil Engineering's 2026-07-30 post — 693 points on Hacker News — enumerates six mechanisms locking agent state to a single provider: reasoning tokens returned as opaque encrypted blobs, web-search passages the model sees but the caller cannot, compacted context only the original provider can decrypt, subagent instructions hidden inside encrypted payloads, server-side response retention on by default (OpenAI 30 days, Gemini 55 days), and encrypted inter-agent messages that make multi-agent auditing impossible. The authors propose seven portability principles, anchored on 'the local event log is canonical,' and call out labs for permitting distillation into their own models while prohibiting it outbound. For anyone running long agent sessions, this is the concrete checklist for whether your transcripts can survive a model switch.
↳ Follow the thread