The specific friction: every provider switch resets context
Alice is the senior PM on Q3 forecast. She spent 3 hours in Claude refining model assumptions, walking through edge cases, building shared vocabulary with the model. Tomorrow morning she's OOO; Bob takes over. Bob's laptop is on the ChatGPT-Team license. He has to either (a) re-brief Claude from scratch for 30 minutes or (b) try to summarise Alice's session in a long ChatGPT prompt and hope nothing important got lost. The math: 30 minutes × every provider switch × every employee transition × every week = a 2-3 FTE drag on a 50-person AI-forward team.
The portable manifest is the fix
GenZAgents stores every Alice-Claude session as receipts. The /v1/agents/[did]/portable endpoint assembles a portable manifest: the conversation's system prompt + a memory snapshot of the working context + a receipt digest of the last 10 receipts. Bob opens ChatGPT, calls the GenZAgents Custom GPT or pastes the manifest into a new ChatGPT chat, and ChatGPT now has Alice's context. No re-briefing. The handoff time goes from 30 minutes to 90 seconds.
How it works under the hood
The MCP tool restore_chat assembles three pieces: (1) the seed system prompt that initiated the original conversation; (2) a memory snapshot — the working assumptions, key decisions, agreed terminology — extracted from the receipt body via a small LLM pass; (3) a digest of the most recent 10 receipts as a "what we just did" recap. The receiver model gets enough context to act as if it had been the original conversational partner. The portable manifest is JSON, model-neutral, and signed by the source agent so the receiver can verify it wasn't tampered with.
Why the LLM vendors won't ship this themselves
Anthropic profits from continuity inside Claude. OpenAI profits from continuity inside ChatGPT. Cross-provider portability damages both. They won't build it; they would never have built Plaid for cross-bank account aggregation either. The infrastructure layer above all providers has to be neutral by structural necessity — same logic as Twilio sitting above AT&T/Verizon. That's where GenZAgents sits.
Edge cases the portable manifest handles
Long sessions (50+ turns) get summarised by the snapshot LLM to fit the receiver's context window. Tool calls that referenced files get path-rewritten if the receiver's working directory differs. Provider-specific features (Gemini grounding, Claude artifacts) get translated to provider-agnostic representations. The receiver doesn't need to know it's a translated session — it just sees a coherent system prompt + memory snapshot and continues working.
Quantified ROI for a 50-person team
Conservative numbers: 50 people × 0.5 provider switches / week / person × 30 minutes saved / switch = 12.5 hours / week recovered. At a £80/hr loaded engineering cost that's £1000/week ≈ £52k/year in recovered time. The Enterprise tier (£499/month = £6k/yr) pays back inside the first month. The remaining 11 months are pure margin — and that's before counting de-dup savings, employee-handover savings, and compliance posture upgrades from the same product.