The Agentic Lab

Companion code for The Agentic Lab: Build AI Agents for Biotech and Pharma R&D, by Suryaprakash Tripathy.

Twelve builds, one folder each, in one repository. Clone it once and you have everything in the book. Each build stands alone: open its folder, run its tests, and nothing else needs to be installed.

Which builds belong to which chapter

Three chapters cover two builds each, so their printed address lands on the list of all twelve rather than on one folder. This table says which two.

The builds

• 01 First Agent and 02 Tool Belt
  An agent loop written against the SDK and nothing else: a step cap, a trace, a budget, an error policy and a write gate. Then the same tool checks done properly, validated at the dispatch boundary. lab.nexagenlabs.com/ch03
• 03 Triage Agent and 04 Dual Screen
  Screening records against criteria somebody wrote down, where every record leaves a verdict or a logged gap. Then two independent screens, their agreement measured rather than assumed, and disagreements adjudicated. lab.nexagenlabs.com/ch04
• 05 Wrangler
  An unfamiliar instrument export turned into a tidy table, with a column mapping a human approves once and six assertions that run every time. lab.nexagenlabs.com/ch05
• 06 Plate Mapper and 07 Protocol Adapter
  A combination matrix with the dilution arithmetic checked before the pipette, and the analysis committed before any data exists. Then a protocol adapted to a different cell line, with a diff naming what changed and what was carried over unchanged. lab.nexagenlabs.com/ch06
• 08 Dock Loop
  Docking treated as a ranking rather than a measurement, with a redocking control and an enrichment control the campaign runs on itself. lab.nexagenlabs.com/ch07
• 09 ELN Bridge
  Read freely from a notebook, write only through an approval gate. The interface has no update and no delete, and every number is checked against the design file before a human sees the diff. lab.nexagenlabs.com/ch08
• 10 Run Manifest
  Content-addressed provenance, and two kinds of replay: one that re-executes and needs the model, one that rebuilds from the stored trace and needs nothing at all. lab.nexagenlabs.com/ch09
• 11 Red Team
  Thirty-one faults planted in the earlier builds on purpose, and an honest count of how many their checks actually catch. lab.nexagenlabs.com/ch10
• Templates (no build)
  Chapter 11 ships paperwork rather than code: the documents that belong around an agent rather than inside it. lab.nexagenlabs.com/ch11
• 12 Repurposing Desk
  The eleven builds connected into one system, from literature to a ranked shortlist, with three human checkpoints and a single manifest behind it. lab.nexagenlabs.com/ch12

Chapters 1 and 2 ship no code, so there is no /ch01 or /ch02. If you typed one, you are in the right place already.

Model names

Model names change faster than books do, so every build reads its name from one configuration file rather than from source. MODELS.md holds the current list with a dated changelog.

References

Appendix D as machine-readable data, with a DOI on every entry that has one and the verification status of each: lab.nexagenlabs.com/references

Errata

Corrections to the printed book are listed on the errata page. If you have found an error in the book or the code, please report it: the page says how.