Agent-Native Research Artifacts

Background

Knowledge moves between people as documents: papers, specs, design reviews, post-mortems. These formats were built for human readers. Increasingly, the reader is an agent, and so is the author.

The pain

Documents pay two taxes. The storytelling tax: real work branches, hits dead ends, and pivots, and the write-up keeps only the winning path. The engineering tax: prose precise enough to convince a reviewer leaves out what someone needs to reproduce the work.

The missing pieces live in notebooks, chat threads, and people's heads. Every new agent that picks up the work pays to rediscover them.

The challenge

The format has to capture the whole branching process with exact specifications and evidence. It cannot ask people to write more. It has to stay checkable by machines, and agents must be able to fork it, rerun it, and compare results against the original.

Our approach

An Agent-Native Research Artifact (ARA) has four layers. The cognitive layer holds the claims, concepts, and problem statement. The physical layer holds the code, configurations, and exact environment. The exploration graph holds every branch, including dead ends and pivots. The evidence layer ties each claim to the raw results behind it.

A live research manager captures the artifact in the background while people and agents work together, so nobody has to package the work afterward.

The outcome

Given the same work as an ARA, agents understood it better and reproduced it more often. The preserved failures helped the next agent most when it tried to extend the work, because knowing what does not work is half of research.

Knowledge becomes something the next agent inherits, the way it inherits code: with version control, a dependency graph, and a record of who changed what.

On silicon

When a block is done, the handoff usually carries the final RTL, the constraints, and a review deck. The failed floorplan scripts, the rejected pipelining experiments, and the exact tool versions stay behind.

Packaged as an artifact, the block would carry all of it. The cognitive layer holds the design intent and its claims. The physical layer holds the RTL, constraints, scripts, and tool versions. The exploration graph holds every option tried, including the rejected ones. The evidence layer links each decision to its reports.

A team working on a derivative chip, or a new agent, could then rerun any branch in the original environment, change one assumption, and compare. Over time, a company’s design history becomes material its self-improving loop can learn from.