Week 12 · lesson

Lab: Audit Multi-Agent Handoffs

Mission

Run the local Multi-Agent Handoff Audit Lab. Compare verified, schema-mismatch, and unauthorized-action traces, inspect four controls, and export a Multi-Agent Handoff Audit Record.

Guided start

Select Verified hold and run the audit.
Confirm the action is held because owner approval is absent.
Select Schema mismatch and find the ambiguous date failure.
Select Unauthorized action and identify the broken approval gate.
Explain why only the final trace is unsafe despite similar messages.

Procedure

Preserve each actor, message, requested action, and outcome.
Audit declared capability, provenance, shared schema, and approval enforcement.
Compare a safe hold caused by missing approval with a failed enforcement.
Identify the accountable human role and recovery action.
Preserve your first agent-consensus claim and write a bounded correction.
Complete Observation, Measurement, Decision, and Limitation.
Export the record and confirm it names ai-for-beginners-multi-agent-handoff-audit-lab.

Success criteria and boundary

every actor, handoff, control, and result remains visible;
schema failure is separated from authority failure;
absent approval causes a hold in the compliant trace;
unsafe execution is named without simulating the external action; and
the limitation rejects claims about fairness, safety, privacy, accountability, or performance outside the supplied traces.

Use only fictional traces. The Lab creates no agents and sends no messages.

Robotnix Lab · contract v1

Multi-Agent Handoff Audit Lab

Audit fictional agent messages for capability, provenance, shared schema, approval enforcement, and accountable outcomes.

Help and boundaries

A deterministic browser trace viewer creates no agents, sends no messages, and performs no external action.

No personal data, private prompts, or API keys. This Lab makes no hidden remote writes and does not persist your work after the page session.

No result creates an agent, sends a message, performs a booking, or authorizes action.

Lab draft is active for this page session only.

Objectives

Trace actors, messages, evidence, schemas, authority, and outcomes.
Distinguish safe holds from schema and approval failures.
Connect inspectable controls to responsible-system accountability without overclaiming.

Procedure

Inspect traceRecord actors, messages, requested actions, authority fields, and outcomes.
Audit contractsCheck declared capability, provenance, and shared schema.
Test authorityVerify absent approval causes a hold and identify failed enforcement.
Bound responsibility claimSeparate trace evidence from fairness, safety, privacy, accountability, and real-system performance.

Workbench

Audit contracts and authority across every handoff.

Trace scenario

Select a supplied trace and run the deterministic handoff audit.

Boundary: These fictional traces create no agents and send no messages. They cannot establish fairness, safety, privacy, accountability, or performance in a real multi-agent system.

Text and static fallback

Every trace, message field, control result, outcome, and boundary appears in accessible text and tables.

Verified and schema-mismatch traces hold; unauthorized trace fails approval enforcement and records unsafe execution.

Evidence record

Record only the evidence requested. Do not enter names, personal information, private prompts, or credentials. This draft stays in memory only unless you explicitly export it.

Handoff traces *Record every actor, message field, requested action, and outcome.Control audit *Record pass or fail evidence for capability, provenance, schema, and approval enforcement.Coordination claim repair *Preserve and repair one agent-agreement-means-responsible claim.Limitation *State what fictional traces cannot establish about a live multi-agent system or affected people.

Success criteria

Actors, message fields, and outcomes remain explicit.
Schema and authority failures are distinguished with evidence.
The correction follows supplied trace evidence only.
The limitation rejects unsupported fairness, safety, privacy, accountability, and performance claims.

Complete every procedure step and required evidence field before export.

Import evidence JSON

Vocabulary lab

Flip the idea, not just the card

Explain the term before you reveal the back. Then compare your explanation with the definition, example, and warning.

Read all terms without animation

capability: A declared operation a component is technically able to perform. Example: The action agent can submit a booking request. Do not confuse it with: Capability is not current authorization.
provenance: Evidence recording where an input or claim came from. Example: A research result includes a source reference. Do not confuse it with: Provenance does not automatically make the source correct.
least privilege: Granting only the minimum access needed for a bounded task. Example: The planner can propose but cannot execute. Do not confuse it with: Convenience is not a reason for broad authority.
fail closed: Blocking action when required evidence or authorization is missing or invalid. Example: An ambiguous date produces HOLD. Do not confuse it with: A silent best guess fails open.
Multi-Agent Handoff Audit Record: Evidence preserving messages, controls, outcomes, claim repair, and limitation. Example: The Lab export. Do not confuse it with: It is not evidence from a live agent system.