Pegasus Agent Mail Case Study

The problem

AI agents on different computers needed to exchange useful work without an always-running service reading every message.

The system needed to move enough context for another agent to continue the task, protect the message from the wake-up service, avoid keeping a model active all the time, and make receipt clear.

Constraints

Identity, exclusivity, recovery, and privacy had to be explicit.

Messages are delivered to an exact task rather than a vague machine or user target.
The coordination layer carries context without interpreting or expanding its authority.
Wake notifications contain only what is needed to locate the intended work.
One recipient claims each item, preventing duplicate handling.
Interrupted work can reconcile, retry, and report a truthful state without retaining private content indefinitely.
Successful pickup removes the bounded mail item rather than retaining it indefinitely.

What it enables

Useful cross-computer handoff without a service reading the message.

Exact deliveryA request reaches the intended task rather than whichever process happens to be listening.

Private wake-upThe signal used to resume work does not need to inspect the substantive message.

Single ownershipOnly one recipient can claim a work item, keeping repeated delivery from becoming repeated action.

Recoverable handoffInterrupted transfers and repeat requests remain diagnosable and can finish in a clean state.

The public explanation stops at the product behavior. Message format, transport, identifiers, machine topology, retry mechanics, access details, and raw diagnostics remain private.

What Shayne owned

The simplifying product rule and its acceptance behavior.

Problem framingDefined the need for cross-computer context exchange without a body-reading relay or resident model loop.

Architecture simplificationSeparated bounded file context from metadata-only notification and required exact task identity.

Lifecycle requirementsRequired deletion after pickup, repeat-request reuse, exclusive consumption, bounded retries, and zero-pending cleanup.

AcceptanceDirected packaged two-computer request/reply, second-request reuse, and final clean-queue verification.

AI role

Implementation assistance within a human-defined trust model.

AI agents contributed implementation, tests, diagnosis, and packaging. Shayne retained the context-versus-notification boundary, identity and deletion rules, simplification decisions, failure interpretation, and acceptance criteria.

Verification

Executable checks plus substantive two-machine use.

Small typed surfaceOnly the bounded coordination operations are exposed.

Bounded message bodiesContext remains size-limited and explicit.

Executable validationPackaging and behavior checks completed cleanly.

2 computersDiscovery, connection, substantive request/reply, repeat request, and clean queue.

Two self-contained executables were packaged. A second independent request reused the same tasks and sessions, and the acceptance ended with zero remaining test mail.

Result & next stage

A working Windows prototype exercised between two computers.

The system was tested, packaged, installed, and substantively live-verified between two Windows computers. Its current scope is exact-task coordination on a controlled local network; broader network and security validation is the next expansion stage.

Current live evidence covers discovery, connection, substantive request/reply, a repeated request, and a clean final queue across two Windows systems.

Move AI work between computers without exposing the message to the wake-up service.