Handoffs with receipts: when the second agent never saw the first one's work
A two-session experiment on dev.to shows a good handoff transfers evidence, not trust: a receipt per action, verifiable results and a clear next step — the same design as Soot's task logs and traces.
A fascinating experiment recently showed up on dev.to: *"The Second Agent Never Saw the First One's Work: A Live Two-Session Zambo Chain."*
The author ran a job across two completely independent sessions. Session A finished its share, wrote receipts for everything it did — including failures and refusals — then packed up a handoff bundle. Session B picked up that bundle and continued the work, without ever replaying the full prior context or re-running earlier tool calls. The two sessions shared no memory, no conversation thread.
The striking part: Session B still continued correctly — because it didn't need to *trust* Session A. It could *verify* what Session A had done.
An abstract "Done" is the enemy of a good handoff
When agents — or honestly, humans — hand work to each other, the usual failures fall into three patterns:
- Dumping the entire context: Forwarding the whole chat history, hundreds of tool calls, tens of thousands of tokens. The receiver has to re-derive everything from scratch — expensive and easy to get lost in.
- Blind session sharing: "Just read my session." The receiver can't tell what actually finished, what was an assumption, and which errors were quietly skipped.
- A bare "Done": The sender claims completion; the receiver has to take it on faith. No evidence, no way to check.
All three fail at the same point: the receiver cannot answer *"What actually happened?"* without digging through everything again.
What a proper handoff package needs
The experiment distills a minimal-but-sufficient handoff into three components:
- A receipt per action: Every operation attempted must leave non-repudiable evidence — what succeeded, what errored, what was denied. Receipts don't decorate results; they prove progress and integrity. A "denied" receipt is worth more than a promise of "almost done."
- A verifiable results summary: A concise conclusion pointing at real artifacts — the file that was written, the row that was inserted — so the receiver can open it and see, not believe.
- A clear next step: What remains unfinished, what was already tried and failed (so it isn't repeated), and exactly where the receiver should start.
The golden rule: whoever takes over — human or another agent — should understand the state of the work from the package alone, without excavating the source context.
In Soot, every handoff carries evidence
Soot is designed around exactly this philosophy: no abstract "Done" — everything leaves a verifiable trail on your machine.
- Tasks with status and logs: Every job assigned to a Soot is a task with an explicit state — doing, blocked, review, done — plus a log of notes: what changed, what's next, why it's blocked. Read the log and you know the progress precisely; no guessing.
- A trace of every tool call: Each tool call — the model's proposal, the runtime's grant decision, the execution result — is recorded in the trace. Even a call denied for missing permissions is preserved as a receipt instead of vanishing silently.
- Real files and data as evidence: When a Soot reports a draft written, the file sits in your workspace folder. When it reports a data row written, you can query that row. The receiver opens the artifact and verifies immediately.
- Report blockers instead of going silent: When a Soot can't proceed — missing permission, a decision only a person can make — it reports a blocker stating exactly what's needed. That, too, is a receipt. No agent quietly stalling while the next person assumes the job is done.
- Review before closure: Work requiring human judgment sits in review, waiting for your sign-off — a proper agent-to-human handoff with its own state and evidence.
This is why, when you audit a chain of work produced by several Soots in sequence — or when a new Soot picks up where a previous one left off — you never rely on the story. You open the trace, the task log, the resulting files, and verify for yourself.
One small Soot, one clear task, clear evidence
This is, once again, why Soot holds to its motto: one small Soot, one clear task, clear evidence.
- One small Soot: a narrow scope with minimal tools — easy to hand off because there's little surface to explain.
- One clear task: measurable completion criteria — that's what makes a receipt meaningful.
- Clear evidence: output as real files, data and traces on your machine — things the receiver can verify instead of accepting a bare "done."
A handoff isn't a transfer of trust. A good handoff is a transfer of evidence.
---
Learn more about task logs, traces and how Soot writes a receipt for every action at trysoot.com/docs. Download Nest to run it on your machine at trysoot.com/download.