Agentic AI

What's New?

June 2026: Perplexity adds "brain" so that it studies the most recent actions related to a topic overnight, just like a dedicated graduate student
June 2026: Google releases Spark, a personal agent

Agentic AI Explained

Medicine is entering a fundamental shift from using artificial intelligence as passive tools to deploying AI as collaborative teammates. Standard large language models (LLMs) operate in a reactive, single-turn capacity—meaning they rely on a human user to provide a prompt, and they produce an isolated text output in return. In contrast, agentic AI systems are semi-autonomous and can independently perceive their environment, create multi-step plans, and use external tools to pursue defined goals without needing human intervention at every step. Importantly, they have memory, which is critical for multitasking. In 2026 we saw a shift in how LLMs work and what they are capable of. As a result, many of the current (frontier) models have agentic qualities:

Anthropic Claude (e.g., Opus 4.6 / Sonnet 4.6): Its tendency to flag ambiguity makes it highly trusted for autonomous, multi-step workflows with less oversight.
OpenAI GPT (e.g., ChatGPT Thinking / GPT-5.x): Excellent at complex multi-step planning, tool orchestration throughput, and multimodal processing.
Google Gemini (e.g., Gemini 3.0 Pro) is known for its massive context window and robust performance in structured, multi-step agent environments.
Kimi 2.6 has an "agent swarm" option so it can assign multiple subagents depending on the complexity of the tasks

Other agentic systems worth mentioning

Manus Ai: The first public agentic system offered in mid-2025
Microsoft Copilot: uses a combination of LLMs
Perplexity Computer: offers single LLMs or an agentic approach
Claude Cowork: uses agents to accomplish multiple tasks

Despite the appearance of agentic systems, most of US medicine remains stuck in the chatbot era. We aim primarily to get an expert unimodal answer from a single prompt. Contrast that with an agentic system that can continuously synthesize information, execute complex reasoning chains, and adjust its actions based on new data, rather than merely functioning as a sophisticated calculator or an advanced search engine. For instance, while a standard LLM might simply draft an email when asked, agentic AI can independently decide to search a medical database, retrieve patient lab results, recognize an anomaly, and formulate a care plan based on those findings before the clinician even logs in. Figure 1 shows a standard agentic AI workflow.

Figure 1: Agentic AI workflow (nano banana)

How Agentic AI Can Be Used in Medicine

In healthcare, agentic AI has the potential to automate complex, multi-step administrative and clinical workflows. Some of the most promising applications include:

Comprehensive Chart Reviews and Triage: Agentic AI can act as an assistant that reads a patient's entire electronic chart, including recent lab results, physician notes, and histories, to suggest possible differential diagnoses or flag missing preventative care, such as overdue mammograms or colonoscopies.
Clinical Workflow Orchestration: Unlike traditional tools, agents can communicate with multiple hospital systems to execute real-world tasks. For example, if a patient needs an MRI, an agent could automatically analyze clinical notes using natural language processing, coordinate with the scheduling system, and cross-reference the patient's record for safety contraindications (like a pacemaker) to preempt any risks.
Automating Medical Documentation: Systems are already being used to process audio from doctor-patient interactions to automatically generate and format discharge summaries, referral letters, and clinical notes directly into the electronic health record (EHR). In real-world trials. Agents could summarize longitudinal care into a useful summary.
Revenue Cycle and Patient Engagement: Agents can autonomously detect coding errors, submit insurance claim appeals, and act as virtual assistants to triage patient symptoms over the phone or via chat.

Agentic AI System Examples

Enterprise-level Agentic AI

Several advanced models and platforms are currently being adapted or explicitly built for healthcare applications:

Google's Med-Gemini processes medical text and medical images, including X-rays and 3D CT/MRI scans.
--Autonomous Web Searching & Grounding: It actively searches the internet for up-to-date medical research, evidence-based guidelines, and clinical data
--Multi-Step Reasoning: Unlike a basic chat model, Med-Gemini utilizes techniques like uncertainty-guided search, chain-of-reasoning, and subtasking. It evaluates its confidence and takes further actions if necessary.
--Tool Orchestration: It can connect directly with external tools, like the electronic health record (EHR) retrieval systems, to perform "needle-in-a-haystack" analyses across thousands of pages of patient data.
--Multimodal Capability: It natively analyzes not just text but also medical images (like X-rays and MRIs), genomic data, and videos.
Medical Intelligence for Reasoning and Action (MIRA) is a recently reported German research project that used MIMIC-IV EHR data
--Two agents: A patient agent and the medical agent (MIRA)
--Performed medical tasks: MIRA queried the history and physical exam and could order labs, blood cultures, scans, medications, procedures, and surgery. It could also perform triage for hospital admission.
--Outperformed physicians: MIRA outperformed physicians in diagnostic accuracy. Decisions made followed guidelines and were safe and appropriate.
Articulate Medical Intelligence Explorer (AMIE) is the longitudinal assessment of outpatient treatment by Google researchers.
--Two agents: Also has two agents:
--Study design: AMIE results were compared to 21 primary care physicians (PCPs) on 100 multi-visit cases designed to match NICE and BMJ Best Practice guidelines
--Ensemble refinement: took 4 different treatment plans developed and came up with a consensus, mimicking a real medical treatment board
--AMIE superiority: Performance was superior in multiple areas.

Med-Gemini. https://research.google/blog/advancing-medical-ai-with-med-gemini/
MIRA. Towards Autonomous Medical Intelligence Agents. https://www.nature.com/articles/s41586-026-10675-5
AMIE. Towards Conversational AI for Disease Management. https://www.nature.com/articles/s41586-026-10764-5

Personal Agentic AI

A personal agent makes your actions more efficient by embedding itself into your computer. It is primarily an administrative assistant. There are many personal agents available today:

OpenClaw and its many clones. Despite the program being the first example of a personal agent, it has security vulnerabilities. The software is open-source, but you must sign up for an LLM API service. The concern is that this model is "always on," so there is the potential for it to run up unexpected charges while you sleep
Perplexity Computer is another example of a personal agentic AI. In June 2026 it released "Brain," which is a self-improving memory system. It builds a context graph of the work Perplexity Computer does. Brain periodically, particularly overnight, reviews the context graph and teaches itself how to improve.
The more work you do, the better and more efficient Brain becomes. There is some data to show objective improvement in accuracy and recall
Google Spark was released in June 2026 but is available for testing only by Ultra subscribers. It will assist the Google Workspace with tasks related to Gmail, Google calendar, Google Drive, etc.
Claude Cowork

We will focus on Claude Cowork because of its maturity and association with other Claude programs.

Claude Cowork

Cowork is Anthropic's desktop AI tool for non-technical knowledge workers. Unlike the chat interface, it handles full multi-step tasks autonomously—you give it a goal, it works across your local files and applications, and it returns a deliverable. It's currently available on all paid plans via the Claude desktop app. The LLMs that are currently available as of June 2026 are Sonnet 4.6, Opus 4.6, 4.7, and 4.8. Claude 5 (Fable) is currently unavailable. There are two modes: "Ask before acting" and "Act without asking." Another mode is called effort: low, medium, high, max, and thinking.

Core Features

File and document work—organizes, renames, deduplicates, and sorts folders; synthesizes multiple source documents into structured drafts; extracts structured data from contracts, records, and reports.
Research synthesis — reads across a set of sources and returns a summary ready for review, rather than making you coordinate the search yourself.
Scheduled tasks — can run jobs automatically on a schedule (e.g., a weekly news digest like the one just generated).
Plugins and skills — installable bundles that extend Cowork's capabilities with specialized tools and workflows.
Artifacts—create persistent, live HTML pages that pull fresh data from connected tools each time you open them (useful for recurring dashboards or trackers).

Integrations

Cowork sits alongside Claude Code (developer-focused), Claude for Chrome (browser agent), Claude for Slack, and Claude for Microsoft 365. They share the same underlying models and can connect to the same external services via plugins. Here are some companion programs worth knowing about

Claude projects: Are self-contained workspaces within the Claude chat interface where you can upload documents, set custom instructions, and maintain a persistent knowledge base across multiple conversations. Instead of re-explaining context each time you start a new chat, a Project remembers your uploaded files and preferences, making it ideal for ongoing work like a research topic, a recurring report, or a client engagement. On paid plans, Projects use RAG to scale well beyond normal context limits, and on Team/Enterprise plans they can be shared across an organization.
Claude skills: These are reusable tasks. For example, I provided Claude with instructions on how I like my slide decks, which makes the results more predictable. I also created a routine task for exploratory data analysis (EDA). Any time I update a new CSV file, I can mention EDA, and it automatically carries out the multiple steps. Skills are composable (multiple can activate in one session), portable across Claude apps and Claude Code, and available from a growing directory of partner-built options from companies like Notion, Figma, and Atlassian. They're available on Pro, Max, Team, and Enterprise plans.
Claude design:
Claude dispatch: Use your smartphone to give Claude Cowork tasks on the fly (beta program)
Claude Code: Is Anthropic's agentic CLI tool that lets developers delegate coding tasks directly from their terminal — writing, debugging, refactoring, and navigating codebases through natural language. It runs on the same underlying models as Claude but operates in a developer-native context, with deep access to file systems, shell commands, and version control.

Features of Interest to Healthcare Workers

Literature and research — the bio-research plugin connects to PubMed, ClinicalTrials.gov, bioRxiv/medRxiv, Consensus, and ChEMBL. You can search the literature, pull trial details, and analyze drug bioactivity without leaving Cowork.
Document preparation — drafting clinical summaries, patient education materials, grant sections, or protocol documents from source files. A reorganization of my downloads folder into 15 logical folders, such as documents, PDFs, images, videos, etc. It organized the content by type and when it was added. Much better than the random chaos that existed
Scheduled briefings — automated weekly or daily reports on medical AI news, literature updates, or trial activity in a specific disease area (like the report generated earlier today). I have a weekly summary of medical AI, as an example.
Data extraction — pulling structured information from dense clinical documents, lab reports, or PDFs.
Presentation creation — the MIS slide template skill you have builds branded PowerPoint decks directly from a topic or outline.
Persistent dashboards — live artifact pages that can display up-to-date trial status, literature counts, or other tracked metrics each time you open them.
MCP Connectors:

The main limitation to keep in mind: Cowork operates on your local files and connected services, but it is not integrated with EHR systems or clinical workflow tools (Epic, Cerner, etc.) out of the box—that would require a custom connector.

Personal Cowork Use Cases

I asked Cowork to organize my download and data science article files into separate folders that represent major topics. Within the folders, the PDFs are organized chronologically
I gave Cowork a synthetic discharge summary and had it do the following:
--Outline the optimal follow-up care for the next year for a patient who was discharged with a myocardial infarction
--Generate a referral letter to the cardiologist for a one-month appointment post-stent placement
--Schedule an appointment to a smoking cessation clinic locally

Challenges, Concerns, and Drawbacks of Agentic AI

While highly promising, the structural autonomy of agentic AI introduces novel and severe safety risks that standard LLMs do not possess.

Hallucinations and Reasoning Failures: Because agentic systems link multiple steps together, a "hallucination" (a fabricated fact, such as an incorrect drug dosage or fake medical citation) generated at step one can silently contaminate the entire reasoning chain. This can result in a deeply flawed, yet highly coherent and convincing, clinical recommendation.
Automation Bias: Because agentic systems present their findings with high syntactic fluency and apparent confidence, clinicians frequently exhibit "automation bias," uncritically deferring to the AI's outputs. This phenomenon suppresses the natural human oversight that is supposed to act as a safety net.
The Accountability Dilemma: Legal and moral accountability frameworks are not equipped for autonomous AI. Clinicians are placed in a "double bind": they are legally liable for patient outcomes but lack the time or technical capacity to verify the AI's opaque reasoning process. They must either spend valuable time manually duplicating the AI's work—negating the tool's efficiency—or trust the agent blindly and absorb the liability risk.
The Workflow and "Deployment" Gap: Despite high technical scores on benchmarks, the industry faces an "autonomy contradiction." Systems heavily marketed as "agentic" operate with sharply limited autonomy in actual practice due to safety and liability constraints, requiring heavy human supervision for nearly every high-stakes diagnostic task. Furthermore, healthcare data is highly fragmented, making it very difficult for AI agents to seamlessly integrate and navigate across disparate hospital systems
The video below summarizes some aspects of agentic AI and healthcare