Portfolio · Project 11 · AI in HEOR · November 2026
From one agent to a team
Project 10 paired two agents: an author that drafts a study protocol and an independent reviewer that checks it. This project adds a third worker in front of them, a literature reviewer that runs a targeted literature review and hands over only verified, PubMed-traced evidence.
You run all three with one command, and they behave like one agent. This page explains why it stays as three, where you still make the calls, and how many review rounds is enough. You can download the team at the end.
Skill → pair → team Three steps up
The rule that never changes: one agent, one skill, one role. Moving up the ladder adds workers and the hand-offs between them. Each skill still works on its own, and improving a skill improves every agent that uses it.
Meet the team
You
Write a one-page brief: the question, the data source, and the literature you need.
Literature reviewer
Runs the targeted review: search, screen, extract and verify.
Evidence pack
Facts and definitions, each traced to a PubMed ID, plus honest gaps.
Author
Drafts the protocol and may cite only the evidence pack.
Reviewer
Checks the protocol and the review, and traces every citation back to the pack. Never sees the author’s notes
Not ready? Back to the author. Up to three rounds.
Why three agents, not one agent with three skills
No marking your own work
An agent that remembers writing the draft reads it through what it meant to say. The reviewer starts blank and sees only the document.
Evidence only from the pack
The author has never seen an abstract, an excluded study or an uncorrected number, so it cannot cite one from memory.
Some things stay hidden
The reviewer never sees the author’s decision log, and it cannot edit the protocol. Neither rule means anything inside a single agent.
A clean memory for each job
A review can involve 400 records. Keeping them out of the drafting and review stops detail from blurring, such as a subgroup’s sample size.
The cost, and when one agent is enough. Three agents take longer and use more of your usage allowance, and work passes between them as files. For a quick internal first draft that you will check yourself, one session that uses the literature review skill and then the protocol skill is fine. The team earns its cost when the output has to stand up to someone else: an HTA body, a journal or your own QC process.
Your three gates What a test run caught at each
The review skill has six checkpoints. In the team they are grouped into three gates where the run stops and waits for you. The steps between gates run on their own and are logged.
Scope and search
You approve the review question, the eligibility table and the PubMed search strings.
Caught in the testPubMed had silently dropped several search phrases. The two affected strings were rewritten and re-run.
The included studies
You confirm the included set, rule on uncertain records and spot-check a sample of exclusions.
Caught in the testA 95% pilot agreement was the model agreeing with itself, not two reviewers. It was logged as exactly that.
The numbers
You verify every extracted number, including that it belongs to the right group, arm and time point.
Caught in the test1 in 8 extracted numbers (12 of 96) were attached to the wrong group, mostly a parent cohort’s sample size. All of the values themselves were correct.
About the test run. The test used an illustrative brief on advanced kidney cancer: 378 unique PubMed records, 31 included studies, and an evidence pack of 60 facts and 27 definitions. In test mode the coordinator cleared the gates itself and labelled every decision, and a separate verifier agent checked the numbers at Gate C. In real use, the run stops and waits for you.
How many rounds is enough?
Open major issues per review round. The team’s first draft started from verified evidence, so it was done in two rounds rather than three. Citations traced rose from 72 of 78 to 78 of 78.
- Round 1 finds most of the real problems: design gaps, claims the evidence cannot support, and citations pinned to the wrong population.
- Round 2 checks the fixes and catches their ripple effects, such as a new endpoint rule that breaks the analysis plan. A revision should never go out unreviewed.
- Round 3 and later turn up mostly minor points. An author that keeps accepting comments bloats the protocol: in Project 10’s test it grew by 85%. A reviewer asked to look again will always find something.
- the reviewer says READY, meaning no critical or major issues are open;
- three rounds have run;
- a round made no progress, which means the two agents disagree on something you need to settle.
Download the team
The download contains the three agents, the coordinator, a study brief with a literature review section, and the targeted literature review skill (1.1.1) from Project 06. The reviewer uses a built-in checklist drawn from public standards, so it works without any other review skill.
- Get the protocol skill. Download it from Project 02 and unzip it into
.claude/skills/. - Copy the
.claudefolder from this ZIP into the same project. Restart Claude Code if this created theagentsfolder. - Connect PubMed through Claude’s connectors.
- Run it: fill in the study brief, add 3–5 seed papers you already know, and type
/protocol-qc-loop. It stops for you at Gates A, B and C.
Read every file before you run it. It is provided as is, without warranty. A targeted review is not a systematic review and cannot serve as HTA clinical evidence. Protocols still need your usual scientific, statistical and compliance review. Licensed under CC BY-NC 4.0 for non-commercial use with credit; see the Terms of use.
Built on
Be curious, not judgmental.
Design inspired by Spencer Gabor and Dieter Rams’s ten principles of good design.
© 2026 Jigar Rajpura. All rights reserved. No part of this site may be reproduced, distributed or adapted without written permission, except as set out in the Terms of use.
DisclaimerPersonal project, not sponsored by or affiliated with any employer (current or former) or any university I attended; views are my own. Built with AI assistance and checked against Anthropic’s Claude Code documentation as of November 2026, but errors may remain and features change. The test results describe one run, in test mode, on an illustrative brief; they are not a validation. Full disclaimer