HandoffLedger
The problem
Production work is now split across agents from different vendors — a research agent hands findings to a coding agent, a support agent hands a case to a billing agent. When the run fails, each vendor's trace ends at its own boundary: nobody can prove what the sender actually sent, what the receiver actually received, or whether the handoff was accepted. Teams reconstruct this from scattered logs, and semantic failures (valid-looking but wrong output passed along) leave no error code at all.
The idea
Why now
Cross-vendor agent fleets just became real rather than theoretical: shared multi-player agent workspaces, encrypted messaging between agent sessions across vendors, and cheap decision models that can gate every handoff in real time all landed in the same week. At the same time, the handoff protocol layer (Google's A2A, now under the Linux Foundation) standardises the wire format but provides no verification, receipts, or replay — the plumbing exists, the accountability layer does not.
What it combines
S.Space (shared multi-player agent workspace — the place handoffs happen) + WorldinMotion (encrypted cross-vendor agent messaging — the channel handoffs travel on) + Microsoft Decision 1.1 (a GA decision model fast and cheap enough to gate every single handoff) + Gemini 4 Neon (frontier model as the judgment layer that scores handoff quality at the gate). Each piece alone is infrastructure; together they make a handoff something that can be verified, accepted or rejected, and replayed — a receipt, not just a message.
MVP
Build first: (1) a receipt schema + Python/JS emitter SDK for one framework (LangGraph) that logs sender, receiver, payload hash, and accept/reject; (2) a decision-gate hook that scores each handoff for completeness/contradiction using a cheap model and blocks on failure; (3) a minimal replay viewer (single HTML page) that renders one run's handoff chain. Deliberately skip: multi-framework support, encryption/signing infrastructure, hosted multi-tenant ledger, alerting, and evals — a local JSONL ledger is enough for the MVP.
How it makes money
The moat
No single vendor can be the neutral cross-vendor handoff layer. LangSmith is built by LangChain and is strongest inside LangGraph; the OpenAI Agents SDK's handoffs live inside OpenAI's stack; each vendor's tracer stops at its own boundary by design. A neutral receipt layer requires every vendor's agents to trust it, and a vendor-owned version would be refused by its competitors — channel conflict, not engineering. The open receipt format plus a cross-vendor dataset of real handoff failures is also a data advantage no single-vendor tool can replicate from inside its own walls.
Distribution
Wedge is the open receipt format plus a free replay viewer: framework maintainers and agent-platform teams adopt the emitter SDK because it costs them nothing and makes their own debugging better. Paid tier sells to platform engineering / SRE teams running multi-agent fleets in production — per-handoff-volume pricing, sold through agent-framework docs and platform marketplaces.
Why it wins
Existing observability tools trace LLM calls inside one framework or one vendor's stack; none of them verify what actually crossed a vendor boundary, and none issue a signed, replayable receipt for a handoff.
Risks
Biggest risk is adoption: if only one framework emits receipts, the ledger is a single-vendor trace with extra steps. The MVP de-risks it by making the emitter a 10-line drop-in for the most popular framework first and proving the replay viewer saves real debugging time on a public multi-agent demo before building the hosted product.
Build it with
- S.SpaceShared agent workspace where handoffs occur.
- WorldinMotionCross-vendor agent messaging channel to instrument.
- Microsoft Decision 1.1: a GA decision model for high-frequency agent micro-decisionsFast cheap decision gate on every handoff.
- Gemini 4 NeonFrontier judgment layer scoring handoff quality.
Repo to start from
handoff-ledger — open receipt format, emitter SDKs, and a replay viewer for verified handoffs between AI agents
Evidence
- Multi-agent systems introduce new challenges in orchestration and observability
- A2A Protocol
- LangSmith pricing
- Langfuse pricing
- Multi-agent handoff loss: Anthropic Fellows research
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.