Field (formerly Idea2PaperLab)

Field began as Idea2PaperLab, a prototype for carrying a research idea through literature review, dataset preparation, coding, experiments, analysis, revision, writing, and criticism. Building that loop exposed a broader problem: an Agent needs more than a workflow graph. It needs a place where context, tools, files, approvals, failures, and long-running work remain coherent over time.

Field is the project that grew out of that realization. It is now a local-first Agent workspace where agents, projects, tasks, tools, and activities share one durable operating environment.

From a pipeline to a workspace

The original research sequence remains a useful built-in workflow:

idea -> literature -> dataset -> coding -> experiment -> analysis
     -> revision -> experiment -> analysis -> writing -> critic

But the system no longer assumes every useful activity fits a single pipeline. Conversational Agents can inspect projects, use tools, create child runs, accept follow-up work, pause for input, and resume from durable events. Scheduled and file-change automations can operate in the same environment while the desktop application is running.

This shift changes the central abstraction. A research stage is one possible activity; it is not the container for all state.

Local control plane

Field combines an Electron desktop shell, a React interface, and a managed FastAPI sidecar. FastAPI owns the central SQLite fact store, while large research artifacts remain in ordinary project directories. Each Agent run executes in its own Python worker process rather than sharing the control-plane process.

That design gives the desktop application one consistent place for Agent state, messages, events, tool calls, approvals, checkpoints, and storage telemetry without hiding project files in an opaque database. SQLite uses migrations, integrity checks, WAL, checkpoints, and online backups because local-first should still mean operationally durable.

Context as evidence, not a transcript trick

Field stores append-only raw Events and immutable Episode summaries with provenance. Before a run, a context compiler selects a continuous recent suffix and relevant long-term material to produce an immutable view for the model.

This avoids treating a chat transcript as both interface and memory system. Summaries can be traced to their source events, Agent-to-Agent handoffs can remain bounded, and model-provider cache behavior stays an optimization rather than the authority on what the Agent remembers.

Tools that fail closed

Running tools on a personal machine makes isolation part of the product contract. Field uses native sandbox helpers for macOS, Linux, and Windows; read-only and workspace-write modes fail closed when the required capability is unavailable. Network access is denied by default, and dangerous actions require matching single-use approvals.

This is a substantial evolution from Idea2PaperLab's earlier, softer isolation model. The goal is not to claim that local Agents are risk-free. It is to make the boundary explicit, testable, and visible to the person whose machine is doing the work.

What remains constant

The name and scope changed, but the original question remains: how can an idea survive the distance between a conversation and completed work? Field's answer is to treat execution history, context, project files, and human approval as first-class facts. Research automation is still one important use case, now supported by a more general and durable foundation.