Edgeless Lab / Agent systems
The Return Path
An interactive study of how agents turn conversation into evidence, coordinated work, and useful feedback.
Owner
Receives
Produces
Proposed reference architecture. Generic roles and illustrative scenarios; no live operational telemetry. Playback follows a selected route. Select Field sheet for a light drawing.
What makes this a system?
Context survives the handoff.
A URL is a source. A message carries purpose. Preserve the thread, attribution, and later corrections so the next agent can understand both.
source → evidence → finding
Work has durable state.
Tasks, dependencies, attempts, and review outcomes belong to an authoritative work graph. A conversation can change the mission without becoming a second task database.
attempt → artifact → review
Learning is broader than execution.
An idea can be strategic, tactical, aesthetic, speculative, or useful only for a moment. Preserve its evidence and validity without forcing every finding into a task.
The return path is part of the design.
An artifact is useful when the requester can open it, understand it, and respond. Delivery should preserve the original subject and feed corrections into the next loop.
Where the hard parts live.
Parallelism needs limits.
More workers compete for workspaces, tools, and budget. Admission, isolated attempts, and explicit integration points make concurrency assessable.
Memory needs provenance.
Retained conclusions can become stale. Keep supporting evidence, uncertainty, and recheck conditions so retrieval does not amplify an old mistake.
Retries need history.
A failed attempt may already have changed something. Recovery must inspect prior effects before repeating an action or spending another budget.
A diagram is a hypothesis.
This is a proposed reference architecture. The walkthroughs illustrate intended contracts; they do not report live traffic, deployment health, or measured performance.