By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
Teresa Torres interviews four members of Medable, a clinical trial platform company, about building Agent Studio, a no-code/low-code platform for configuring and deploying AI agents across the clinical trial lifecycle. The conversation covers two production agents, an ETMF document-classification agent and a CRA monitoring agent, and explains why the team chose a platform approach over one-off builds. It goes into architecture choices such as retrieval strategies, ontology layers, custom MCP connectors with authentication, and context management with sub-agents. It also addresses evaluation and human-in-the-loop design in a GXP-regulated environment. The episode is useful for PMs building AI products in heavily regulated domains.
01Key takeaways
- Build a shared platform for agents when multiple use cases and customer types need the same underlying capabilities.
- Normalize messy domain terminology with an ontology layer before asking agents to reason across multiple data sources.
- Wrap external tool connections in a credentialing and authentication layer so agents never handle raw access directly.
- Use sub-agents and automatic tool filtering to keep context windows small and agent behavior reliable.
- Treat human feedback as a noisy signal, not ground truth, when designing evals.
- In regulated settings, document intent, specification and test evidence for every agent behavior from the start.
02Key sections
- Mission and the two agents
- The guests explain Medable's goal of shortening drug development timelines and describe the ETMF and CRA agents built to automate document classification and clinical data monitoring.
- Why a platform rather than one-off builds
- The team argues that a shared platform with models, skills, knowledge bases, and connectors lets them reuse work across products, services engagements, and self-serve customers.
- Retrieval, ontology and tool design
- They compare RAG approaches at scale and describe a unified ontology layer that maps terminology across 13 clinical systems, plus custom MCPs wrapped in an authentication layer.
- Context management and sub-agents
- To avoid tool bloat and context overload, the agents use sub-agents and automatic tool filtering, keeping each step's context focused.
- Evals, human feedback and compliance
- The team designs golden datasets and production monitoring, treats human feedback as imperfect rather than ground truth, and documents intent, specification and test evidence for regulators.
03From the post
“Listen to this episode on: Spotify | Apple Podcasts What if AI could help reduce the 10-plus years it takes to get a new drug to market? That's the driving ambition behind Medable's agentic platform—and the bet that led them to build Agent Studio. In this episode”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.