AI Agent

AI Agent in Evidence Review and Model Building

An AI agent is a piece of artificial intelligence software, usually built around a large language model, that works through a task in turns, deciding at each turn whether to search, read, run a tool or stop, and feeding what it finds into its next decision. For health economists the term comes up in proposals to automate multi-step jobs such as retrieving and screening studies, extracting cost and utility data for a health economic model and assembling submission dossiers. Published validations so far cover bounded tasks within review pipelines, with human reviewers checking the output, while fully autonomous end-to-end reviews have given variable results.

The observe and act loop

Wang and colleagues open their survey of language model agents with the classic definition by Franklin and Graesser, in which an autonomous agent is a system situated within an environment that senses it and acts on it over time in pursuit of its own agenda. The ISPOR working group on generative AI describes agents as systems that combine a language model with memory, task execution, interaction with external tools or data sources and, at times, collaboration with other agents.

Yao and colleagues, who proposed the ReAct method, give a general setup for an agent interacting with an environment.

$$a_t \sim \pi(a_t \mid c_t), \qquad c_t = (o_1, a_1, \ldots, o_{t-1}, a_{t-1}, o_t)$$

where $t$ is the time step, $o_t$ is the observation the agent receives from its environment at step $t$, $a_t$ is the action it takes, $c_t$ is the context made up of every earlier observation and action plus the current observation, and $\pi$ is the policy that maps a context to an action. In a review pipeline an observation might be a page of a trial report, and an action a new database query or a value written to an extraction sheet.

ReAct then lets the model write its own reasoning as a kind of action.

$$\hat{A} = A \cup L$$

where $A$ is the set of actions that change or query the environment, $L$ is the space of language, and $\hat{A}$ is the augmented action space. A step taken in $L$, which the authors call a thought, has no effect on the environment and returns no observation, but it can break a goal into sub-tasks, pull out the useful part of an observation or change the plan when something fails.

Profile, memory, planning and action

Wang and colleagues sum up much of the published design work in a unified framework of four modules. The profiling module sets the role the agent plays, usually written into the prompt, for example an analyst extracting utility values from trial reports. The memory module stores what the agent has observed, and the survey likens short-term memory to the context window of the model and long-term memory to external vector storage that the agent can query. The planning module breaks a task into sub-goals, either fixed in advance or revised in response to feedback from the environment, from humans or from other models.

The action module turns decisions into outputs. Its action space divides into external tools and the internal knowledge of the language model, and the external tools include application programming interfaces, databases and knowledge bases, and other models. An agent compiling inputs for a cost-effectiveness analysis might run a bibliographic search, query a database of unit costs and run an R script, each as a separate action. A knowledge graph or a concept vocabulary such as the wiki's Health Economics Ontology could serve as a knowledge base of defined terms. The Model Context Protocol is an open protocol for connecting language model applications to external data sources and tools.

Reasoning and acting in turns

In its question answering and fact verification tests, the ReAct agent could search a simple Wikipedia interface, look up a string on the current page or finish with an answer, and each trajectory alternated thought, action and observation. The authors report that this addressed the hallucination and error propagation seen when a model reasons without outside information.

Their error analysis shows the trade-off. In 200 hand-labelled trajectories on the HotpotQA task, hallucinated reasoning or facts made up 56% of failures for chain-of-thought prompting and none for ReAct, but reasoning errors made up 47% of ReAct failures against 16% for chain-of-thought. Searches that returned nothing useful caused 23% of ReAct errors, and one frequent failure was the model repeating its earlier thoughts and actions in a loop. Grounding each step in retrieved text cut invented facts, while the quality of the retrieval became the weak point.

Published uses in evidence synthesis and HTA

The ISPOR working group lists possible agent tasks in health economics and outcomes research. In evidence synthesis, agents can retrieve and summarise research findings to speed up a systematic review, in economic modelling they can extract and organise cost and utility data from real-world evidence, and in dossier development they can collate data and format reports. The group calls agents early in development.

A multi-agent review pipeline in spine surgery

Muthu and colleagues built a pipeline of four agents, for deduplication, title and abstract screening, structured data extraction and verification, and describe it as executed entirely locally for data governance. Two extraction agents processed the same studies independently and their disagreements went to human reviewers, while full-text screening stayed a human task. Against dual independent human reviewers in three spine surgery reviews covering 6,214 records, the extraction agents agreed on 1,446 of 1,550 data fields, with a Cohen's kappa of 0.87.

The verification agent flagged 104 discrepancies, all adjudicated by humans, of which 15 were classed as clinically critical, such as intervention and control cohorts swapped or composite outcomes misread. The authors note that all three reviews came from one research group in one clinical area, so transferability to other reviews, such as a network meta-analysis, is not established.

Bounded tasks rather than whole reviews

Fleurence and Chhatwal, responding to the 2025 HTAi Global Policy Forum report, argue that generative AI performs best when it augments clearly defined, bounded tasks. The strongest evidence, they report, comes from screening and structured data extraction under explicit criteria and predefined fields, while fully autonomous, end-to-end reviews have given far more variable results that are hard to validate. They also raise data governance, since health technology assessment work can involve confidential prices and proprietary models, and in many jurisdictions it remains unclear whether uploading these to a third-party AI tool counts as disclosure.

What NICE expects of evidence produced by an agent

The NICE position statement on the use of AI in evidence generation (ECD11), published on 15 August 2024, applies to any AI method used to produce evidence for NICE and makes no separate provision for agents. Submitting organisations should carry out technical and external validation and present the results. They should declare the use of AI, explain the choice of method and report how it was used, including human input, and they remain accountable for the content of the submission. The statement's wider scope and listed uses are covered under artificial intelligence in health.

Recording what an agent did

A model input drawn by an agent needs a record of its origin and of who checked it, and Muthu and colleagues kept audit trails of the origin of each data element, the agent responsible and any human intervention. The W3C provenance ontology, PROV-O, offers a standard vocabulary for such records, in which an agent is something that bears some form of responsibility for an activity, for the existence of an entity or for another agent's activity. Its subclasses of agent include people, organisations and software agents, the last defined simply as running software.

PROV-O also handles delegation. An extraction run can be associated with the software agent that carried it out, the resulting utility value can be attributed to that agent, and the software agent can be recorded as having acted on behalf of the analyst, who also bears responsibility for the result.

Worked example with step errors in an extraction run

The numbers in this example are hypothetical and illustrative, apart from the final step, which recomputes figures published by Muthu and colleagues. Suppose an agent extracts 20 inputs for a cost-effectiveness model, such as unit costs, utility values and transition probabilities, and each input is correct with probability 0.95, independently of the others.

Step 1. The chance that every input is correct is

$$P_{\text{all}} = p^{k} = 0.95^{20} = 0.3585$$

where $P_{\text{all}}$ is the probability that the run contains no error, $p$ is the probability that one input is correct and $k$ is the number of inputs. The expected number of wrong inputs is $k(1-p)$, which here is 1, so a model populated by a single unchecked run would usually carry at least one error.

Step 2. Suppose a second, independent extraction followed by human adjudication of disagreements catches 90% of errors. The residual error rate per input becomes

$$r = (1-p)(1-d) = 0.05 \times 0.10 = 0.005$$

where $r$ is the probability that an input is wrong after checking and $d$ is the share of errors the check detects. The chance of a clean run is then

$$P_{\text{all}} = (1-r)^{k} = 0.995^{20} = 0.9046$$

where $P_{\text{all}}$, $r$ and $k$ are as defined above, and the expected number of wrong inputs, $k r$, falls to 0.1.

Step 3. If both extractions use the same model and prompt, they may make the same mistake on the same field, so their agreement hides the error and the detection rate falls below 90%.

Step 4. In the published spine surgery pipeline, the concordance between agents and human reviewers was

$$s = \frac{1446}{1550} = 0.9329$$

where $s$ is the share of data fields on which the agents agreed with the human reviewers, matching the 93% concordance reported. The 104 flagged discrepancies split into 89 minor and 15 critical, so about one flagged item in seven could have changed a conclusion had it gone unchecked.

Design choices that change the output

Several settings sit between a source document and the values an agent returns, and a reviewer needs them to repeat the work. Muthu and colleagues reported two of them.

• Model and version. They named the model release, set the temperature to 0.2 and capped output at 2,048 tokens, and warned that updates to proprietary software could change results.
• Independent passes. Screening decisions were run twice and accepted only when both passes matched, and extraction used two agents whose disagreements went to people.

Prompts are a further choice, since Wang and colleagues note that minor changes to a prompt can give substantially different outcomes, and an agent depends on prompts for every module.

AI agents beside other agents in health economics

Health economists already use the word agent in other senses. An agent-based model simulates many individuals that follow rules and interact, so that outcomes emerge from their behaviour. The principal-agent problem concerns a person or organisation acting for another with different information or interests, and an AI agent acting for an analyst raises a similar question of who answers for its actions.

Retrieval-augmented generation adds retrieved passages to a model's input, and an agent may use retrieval as one action among several. A large language model on its own produces text, and becomes an agent only inside a loop that lets it act and observe the results.

Limits and common misreadings

Most errors come from reading an agent's output as checked evidence. Four recur in review and modelling work.

1. Assuming tools remove hallucination. Wang and colleagues describe hallucination, the tendency to produce false information with high confidence, as a concern for agents as well as for language models.
2. Ignoring the number of model calls. An agent may query its language model several times for each action, for example to read memory and to plan, so long tasks run slowly.
3. Treating one validation as general. Results from three reviews in one specialty by one group do not give the error rate on a new review question.
4. Equating autonomy with accountability. Under ECD11 the submitting organisation remains accountable for the content, so an agent's output enters a submission only through people who have checked it.

Sources

• Fleurence RL, Chhatwal J. Trust, validation, and accountability: implementing uses of generative artificial intelligence into Health Technology Assessment practice. International Journal of Technology Assessment in Health Care. 2026;42(1):e41. https://doi.org/10.1017/S0266462326103638
• Fleurence RL, Wang X, Bian J, Higashi MK, Ayer T, Xu H, Dawoud D, Chhatwal J. A taxonomy of generative artificial intelligence in health economics and outcomes research: an ISPOR working group report. Value in Health. 2025;28(11):1601-1610. https://doi.org/10.1016/j.jval.2025.04.2167 (preprint read, arXiv:2410.20204v2)
• Lebo T, Sahoo S, McGuinness D, eds. PROV-O: The PROV Ontology. W3C Recommendation, 30 April 2013. https://www.w3.org/TR/prov-o/
• Model Context Protocol. Specification, version 2026-07-28. https://modelcontextprotocol.io/specification/2026-07-28
• Muthu S, Kolarpatti Ponnusamy DV, Viswanathan VK, Rajappan Chandra SK, Sharun K. A locally executed agentic artificial intelligence framework for deduplication, screening, and structured data extraction in spine surgery systematic reviews. JBJS Open Access. 2026;11(3):e26.00102. https://doi.org/10.2106/JBJS.OA.26.00102
• National Institute for Health and Care Excellence. Use of AI in evidence generation: NICE position statement (ECD11). Corporate document, published 15 August 2024. https://www.nice.org.uk/corporate/ecd11
• Wang L, Ma C, Feng X, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science. 2024;18(6):186345. https://doi.org/10.1007/s11704-024-40231-1 (arXiv:2308.11432v7)
• Yao S, Zhao J, Yu D, Du N, Shafran I, Narasimhan K, Cao Y. ReAct: synergizing reasoning and acting in language models. International Conference on Learning Representations (ICLR 2023). arXiv:2210.03629v3. https://arxiv.org/abs/2210.03629