Skip to main content

Research Workflow

A path from reading and questions to experiments, evidence, and ongoing project work.

I use linked notebooks and a set of optional tools to connect paper reading with experiments and claims. Reading, planning, execution, review, and monitoring overlap: a result can send me back to a paper or a new question. Agents help with selected activities; people decide what to investigate and what to claim.

Research Notebook sets out the conventions for linking questions, plans, experiments, and claims.

  1. Activity 1

    Reading

    Reading supplies background and candidate explanations for a research question. Reading Relay finds papers in Markdown lab notebooks and selected PDF downloads, then uses a configurable coding-agent harness to curate a small, persistent reading queue. Reading decisions stay in its local ledger; the notebooks remain unchanged.

  2. Activity 2

    Questions, plans, and experiments

    Record the question, plan the comparison, and declare what evidence would support the claim before running the experiment. Research Notebook links these records. A research-spec block declares evidence sources, quantities to estimate, checks that must pass, and exposure to data, so the tool can validate the design and derive its checks.

  3. Activity 3

    Execution

    Weft runs jobs such as experiments on a workstation, shared server, or rented cloud GPU. It copies the source tree to the selected machine and keeps each job’s logs and outputs. research-site hands a project between a laptop and a remote machine so only one writes to it at a time, and an experiment plan can run where CPU is available.

  4. Activity 4

    Review and evidence

    Compare the outputs with the planned checks, record what the evidence supports, and decide which claims or follow-up experiments are warranted. agent-review records independent reviews with the version examined, evidence provided, and findings. code-provenance checks whether compared results came from equivalent code and dependency declarations.

  5. Activity 5

    Monitoring

    Research Rounds is a macOS app for making rounds of research projects. One window shows each active project’s plans, experiments, and the agent sessions working on it. Weft Status is a macOS menu bar companion and floating HUD showing Weft job, rental, and autopilot activity.

Tools at a glance

  • Reading Relay is a published CLI for curating a persistent reading queue.
  • Weft is a published job runner for local and remote work; experiments are one kind of job it can run.
  • research-spec unpublished validates experiment-design declarations about evidence sources, estimates, gates, and data exposure.
  • research-site unpublished manages a project's handoff between laptop and remote machine.
  • agent-review unpublished records independent reviews of research results alongside reviews of code and documents.
  • code-provenance unpublished checks whether numerical results came from equivalent code and dependencies.
  • Research Rounds unpublished provides a macOS overview of active project plans, experiments, and agent sessions.
  • Weft Status is a published macOS companion for viewing Weft jobs, rentals, and autopilot activity.

Coding-agent launch, command guards, session coordination, reviews, and remote execution are described in Coding Agents. Those tools support parts of the research workflow without defining its required steps.