Agents | Will Bolton

Adversarial Fast-Moving Real-World Domains as Test Beds for Benchmarking AI Scientist Capabilities

Sat, 11 Jul 2026 00:00:00 +0000

ICML 2026: workshop papers and AI for Science meet-up

Fri, 10 Jul 2026 00:00:00 +0000

At ICML 2026 in Seoul I presented:

Learning Clinical-Trial Strategy: Offline Policy Training for Decision Agents: spotlight at the Workshop on Generative and Agentic AI for Biology (GenBio), and poster at the Workshop on Decision-Making from Offline Datasets to Online Adaptation. Paper · Poster
Adversarial Fast-Moving Real-World Domains as Test Beds for Benchmarking AI Scientist Capabilities: Workshop on AI for Science. Paper · Poster

Alongside the conference, the Encode Fellows and I organised and hosted Startups for AI x Science, a meet-up for founders, researchers and builders attending ICML.

Learning Clinical-Trial Strategy: Offline Policy Training for Decision Agents

Fri, 10 Jul 2026 00:00:00 +0000

A2A Net

Wed, 01 Jul 2026 00:00:00 +0000

A2A Net lets AI agents chat and collaborate in real time. Any agent can start a temporary workspace, invite others through a link or a published endpoint, and work together in a hosted environment that provides private and shared spaces, compute, and full traceability. The aim is infrastructure for multiplayer agentic work, particularly across organisational boundaries: the humans involved oversee progress and approve any sensitive decisions, while their agents do the work.

The idea came from personal experience of researching and building with collaborators in 2026: our agents could not collaborate with each other in the way we could, and existing collaborative tools were not designed for a world where everyone has an agent.

The design follows from three observations: agents are session-based, so collaboration needs to happen in real time rather than asynchronously; they move information far faster than people, so they need their own sandboxed compute; and the hard problem is collaboration across institutions, which calls for a neutral, permissioned meeting space.

Built with Python, React, Docker and cloud compute, supporting both MCP and the A2A protocol. Launched in July 2026.

et al.

Tue, 12 May 2026 00:00:00 +0000

et al. starts as a personalised discovery feed and self-improving knowledge base for researchers and their agents. Researchers connect their papers, notes and research tools; et al. indexes this knowledge and turns it into working memory for agents. In return, the agents help the researcher reason across papers, generate citations in any journal format, and surface relevant new literature, hypotheses and experimental methods.

Users can import an existing library or save papers, datasets and code repositories as they browse, with a Chrome extension that captures any webpage in one click. They can then read these materials in the app, ask the agent questions about a single paper or their entire library, and receive suggestions of new work based on what they are reading. Every answer is grounded in the user’s own sources, so each claim can be traced back to the passage it came from.

Behind this sits a metadata ingestion pipeline that resolves records across multiple sources and extracts equations from LaTeX, with OCR and vision fallbacks for scanned documents. Citations can be inserted directly into Google Docs, Word and Overleaf through the Word add-in and browser extension, and a remote MCP server lets an external coding agent query a researcher’s library directly.

I also built a reference checker that verifies whether the citations in a paper are correct and actually support the claims made about them: it parses the bibliography, infers the citation style, resolves each reference and checks every claim against the original source.

By capturing what researchers read, save, ask and eventually test, the platform builds a data flywheel: it improves each user’s view of the scientific frontier while gathering the workflow traces needed to train and evaluate scientific agents. Launched in May 2026 and publicly accessible, with around 2,000 people trying the product.

Dementia Gap Map

Wed, 01 Apr 2026 00:00:00 +0000

Dementia Gap Map is a semantic map of dementia and GWAS research drawn from PubMed, covering several thousand papers grouped by disease area, from Alzheimer’s disease and vascular dementia to Parkinson’s, Huntington’s and prion disease. Each paper is linked to its gene targets, pathways and clinical trials, and connections on the map are based on citations.

Users can draw a region on the map to see the attributes of that group of papers as a feed, or ask the map questions directly, such as which mechanisms have active clinical development and which have stalled, or which pathways have strong genetics but little clinical translation. Built as a demo for a research group in the UK Dementia Research Institute (UK DRI).

AI Scientists Need a Social Network

Thu, 19 Mar 2026 00:00:00 +0000

AGE-nt: intervention intelligence for aging

Sat, 28 Feb 2026 00:00:00 +0000

AGE-nt aggregates data from 11 sources (PubMed, clinical trials, grants, patents and more), normalises the documents into a unified schema, and exposes reasoning tools for evidence grading, trajectory analysis and gap identification through an MCP server, so any LLM client can connect to it.

AGE-nt was built in 24 hours for the Berlin Bio × AI Hackathon (27–28 February 2026), organised by Nucleate Germany, and was selected for the final pitch showcase.

The core idea is that the data layer is the product: agents can only reason well about ageing interventions if the evidence from papers, trials, grants and patents is retrieved, normalised and graded in a consistent way. AGE-nt provides that layer together with tools for grading evidence, tracking how an intervention’s evidence base has evolved over time, and identifying gaps.

Built with Python, FastAPI, FastMCP and a React/TypeScript frontend.

Berlin Bio × AI Hackathon finalist: AGE-nt

Fri, 27 Feb 2026 00:00:00 +0000

A 24-hour hackathon (27–28 February 2026) on real problems in protein design, genome modelling and agentic AI, organised by Nucleate Germany, Project Europe and UNITE, with challenge partners including Google DeepMind. Our team built AGE-nt and was selected for the final pitch showcase, where I presented how agents can retrieve, synthesise and reason over fragmented scientific evidence on ageing interventions.

AI Assistant to Find and Retrieve Patient Cohorts

Tue, 01 Jul 2025 00:00:00 +0000

AI developers need large volumes of specific, high-quality medical data to train and validate their models and to get them through regulatory approval. Hospitals hold that data, but often do not know exactly what they have or how good it is, so finding the right patient cohort turns into lengthy, expensive and opaque negotiations that serve neither side well.

This system tackles that problem with an LLM-driven agent that searches across a hospital’s information systems, including imaging archives (PACS), electronic medical records and unstructured free-text reports, from a single natural-language query. A request such as “female, BMI under 25, more than two scans, subdural intracranial haemorrhage” is decomposed by a query planner into iterative DICOM and FHIR searches, grounded in clinical vocabularies such as SNOMED CT. The agent reconciles the results, validates every patient against the full set of constraints, and returns the cohort together with information on the quantity and quality of the available data.

Once suitable data is identified, developers can request access. Approved requests are extracted, de-identified and shared securely, with the healthcare provider keeping oversight of its data through a CRM-style view of requests. The aim is to cut the workload on hospital admin and IT teams while letting institutions make proper use of their data assets. The initial focus is radiology, where most medical AI development is concentrated, with plans to expand to other types of medical data.

A working prototype built on a hybrid synthetic and public dataset was used by potential customers, and a partnership was secured with an imaging informatics company.