Run a Human-Steered Research Pass
This tutorial shows how to run research as a controlled conversation with an Isomer-equipped agent. The task is to choose a bounded pass, inspect what the agent produced, then continue, redirect, pause, or demand stronger evidence.
Starting Point
You have a prepared Topic Workspace and a next research objective. In the running example, the user repeatedly asked for empirical passes, hypothesis passes, analysis passes, and paper updates, then corrected the agent when evidence or presentation was weak.
Task
Prompt the agent with the research move you want, or explicitly invoke the DeepSci pipeline skill when you know the route. The agent should route to the appropriate DeepSci skill or pipeline, report stage status, produce durable records or artifacts, and tell you what next decision is needed.
Example Interaction
User Action:
$isomer-deepsci-pipeline list passes
AI:
Listed available passes:
empirical-pass,hypothesis-pass,paper-pass,revision-pass,polish-pass, andsubmission-pass.No pass was executed.
User Prompt:
Do an empirical pass using the topic workspace operator workspace.
AI:
Started an empirical pass.
Completed scout, baseline, and idea stages.
Experiment is running in the background.
Next: interpret the experiment result and produce a pipeline terminal report.
User Prompt:
What test has been done?
AI:
Separated project unit tests from research-model tests.
Reported predictor metrics on validation configurations.
Stated clearly that the current model had not yet been tested against real B200 silicon.
User Prompt:
Continue to do experiment.
AI:
Started a bottleneck-refinement hypothesis pass.
Named the hypothesis, target metrics, and expensive run stage.
Reported when the pass completed, which records were created, and what next analysis route was recommended.
Recover Missing Prerequisites
An ordinary task request does not authorize the agent to create every missing input automatically. If you ask for a pass whose accepted prerequisites are missing, the agent should pause before prerequisite mutation, identify the missing records and their producer routes, recommend an order, and keep the original target as the resume point.
User Prompt:
Write the reviewed paper from this topic.
AI:
The paper target is paused before writing because the current topic has no accepted analysis finding. The recommended route is to complete the pending experiment analysis, then resume
paper-passatoutline.You can ask me to run to the paper target, execute only the next prerequisite, inspect or choose another route, or stop without changing prerequisites.
Choose “run to the paper target” when you want the agent to automate routine in-scope prerequisite work and then perform the original task. Choose “execute only the next prerequisite” when you want to inspect each bounded result. The alternate-route choice lets you change the dependency path, while stop leaves the target paused.
Run-to authorization is prompt-scoped and target-scoped. It is not a CLI command, global yes-to-all setting, Project preference, or Run-level Control Mode. The agent keeps separate focused-skill or procedure Runs and terminal reports, refreshes accepted state between them, and stops after the named target. Even with run-to authorization, it pauses for human Gates, material scientific choices, destructive or irreversible actions, credentials or restricted data, material license decisions, unexpected cost or resource use, public exposure, publication, submission, and external side effects.
How to Steer
Research passes work best when you keep the next move narrow. Ask for one modeling pass, one experiment, one analysis, or one paper revision. After the agent responds, inspect the artifacts and decide whether the next prompt should continue the same line or change direction. When a clear target has several routine prerequisites, explicitly ask the agent to run to that target instead of manually acknowledging each intermediate terminal report.
Useful steering prompts:
User Action:
$isomer-deepsci-pipeline hypothesis-pass
User Prompt:
Summarize the generated records and artifacts. Which claim do they support, and what evidence is still missing?
User Prompt:
Pause before the expensive run stage and ask for approval.
User Prompt:
This result points to a modeling gap. Route to analysis and propose the next experiment.
Under the Hood
The agent may use isomer-op-entrypoint, isomer-deepsci-pipeline, or individual DeepSci skills such as isomer-deepsci-scout, isomer-deepsci-idea, isomer-deepsci-experiment, and isomer-deepsci-analysis. Those skills use isomer-cli internally to resolve latest context, write records, attach lineage, and render readable views.
What Good Looks Like
A good pass ends with a concrete artifact, record id, evidence class, blocker, or next route. It should not end with only "I worked on it" or "the model improved."
Common Pitfalls
Do not let one prompt turn into an uncontrolled research spree. Do not accept progress without artifact locations. Do not treat a failed pass as wasted work; the failure often identifies the next best prompt.
Next Tutorial
Continue to Validate With Real Evidence.