A SciOS Program
Automate the tedious.
Amplify your capabilities.
Make more room for discovery.
A 4-week intensive built around your research. Learn to use LLMs by building working agent systems from your own data, processes, and problems.
Applications close Wednesday, August 19. Everyone hears back by Friday, August 21.
You bring your real problems. Grant writing, data pipelines, lab processes, software needs, anything keeping you from your research. Over four weeks, you build the system that solves them, one component at a time.
Knowledge base — full context on your lab, your data, and your work, organized for agents to use
Skills and agents — your repeatable processes encoded as one-command automations, and delegated workers that run them in their own workspace
Memory — what the system learns about your preferences and decisions, and keeps
Orchestrator — the chain that runs multi-step work, with human checkpoints where the work stops for you
Adversarial reviewer — an agent whose only job is to attack the work before you trust it
Improvement loop — the habit that makes the system better every week you run it
Cohort 2 opens with a complete system running, so you see the destination before you build a single piece. This is the Ashworth Group's survey pipeline, demonstrated live on day one, from raw export to reviewed report.
ValidateSkill
A wave of the Ashworth Group's mobility survey lands, six metro sites of numbers and free text, with failures planted throughout. A validation skill checks the structure and halts the run on anything malformed.
Quality controlAgent
Speeders, straightliners, duplicates, out-of-range values, junk text. Every planted failure gets flagged with the rule that caught it, and one under-fielded site puts the whole wave on FLAG.
CheckpointYou
The orchestrator stops and asks. Nothing gets cleaned until you have read the QC report and approved it.
Clean + summarizeAgents ×6
A cleaning agent drops exclusions without ever touching the raw export. Then six summary agents run in parallel, each computing weighted results for its own site, none reading another's output.
CodingAgent
A coding agent reads the group's codebook in full, then codes every open-text answer against it. Where nothing fits it writes UNCODED. It never invents a code.
ReliabilityAgent
An isolated agent re-codes a sample without ever seeing the first pass, and Cohen's kappa measures agreement per question. Below threshold, the pipeline halts.
CheckpointYou
Kappa and every coding disagreement land on your desk before a single figure gets drawn.
DraftAgents
A figure agent plots the results. A writer agent drafts the findings and has no network access, so it cannot invent a number it wishes it had.
ReviewAgent
It reads the audit trail before the outputs, then returns APPROVE, REVISE, or REJECT. On this frozen run it returns REVISE, with three findings that hold up.
The reportOrchestrator + you
Only after the verdict does the report assemble. The run ends where it started, with your judgment.
The Ashworth Group is a fictional policy research center built for this course. Its data is synthetic and its failures are planted on purpose. The pipeline, the agents, the statistics, and the reviewer's findings are real.
The complete system, end to end. Open it full screen to read every step.
Every session ends with something you built and can use right away. You build it alongside a cohort from different fields, and their workflows surface techniques you'd never find working alone.
Your Knowledge Base
You'll Build
A knowledge base that gives AI agents full context on your lab, your research, and your work, started from a copied kit with the Ashworth pipeline inside
You'll Learn
How LLMs work, how the complete pipeline fits together, and how to structure your work so agents can find what they need
Specs, Skills & Agents
You'll Build
A spec precise enough for agents to execute, and your first skills running as one-command automations
You'll Learn
How to write a spec that survives contact with an agent, how to encode your processes as skills, and when to delegate instead of doing it live
Harnesses
You'll Build
An orchestrated pipeline with human checkpoints, an adversarial reviewer, and memory
You'll Learn
How to chain skills and agents into multi-step systems, where your judgment has to stay in the loop, and how to make verification a component instead of a vibe
Ship
You'll Build
Your complete system running a narrow slice of your real workflow, demonstrated live to the cohort
You'll Learn
How to finish a narrow slice completely, and how the improvement loop keeps the system improving after the course ends
Community
After the intensive, you join a community SciOS maintains and runs. Monthly calls, shared resources, and deep dives into verticals like Data, Writing, Knowledge Management, Literature Review, and Software. The 4 weeks end. The learning doesn't.
PIs, postdocs, PhD students. Anyone doing the research and spending too much time on the work around it.
Research engineers, data scientists, librarians. The people building the software, pipelines, and systems research depends on.
Administrators, compliance officers, grants managers. The people who keep institutions running.
If your work supports research and you want to use LLMs well, this is for you.
Cohort 1 ran in April and May 2026 with 14 researchers and research professionals from SDSC, PNNL, Stanford, NCAR, Trinity College Dublin, UCSD, and elsewhere. Cohort 2 keeps what worked and tightens the rest. Skills and agents now share one session, and day one opens with a complete system running instead of building toward one.
The tooling moved between the two cohorts, and it will keep moving. PICoding teaches what holds. Knowledge bases, precise specs, delegation with verification, and your judgment at the checkpoints carry across every model release.
Four Wednesdays: August 26, September 2, 9, and 16. Live sessions run 12:00–1:30 PM Eastern, and you'll need real build time between them. Free, capped at 10. Applications close Wednesday, August 19.
