Thoughtworks Engineer Reveals AI Agent Lessons for Pharma

Building Reliable AI Agents for High-Stakes Industries

Deep research agents represent a significant leap forward in artificial intelligence, moving beyond simple chat interfaces to complex, multi-step reasoning systems. At the 2026 Arc of AI conference, Sarang Kulkarni from Thoughtworks detailed the practical challenges of deploying these systems in critical sectors like healthcare and pharmaceuticals. His presentation highlighted that while tools like OpenAI's Deep Research and Gemini Deep Research Agent offer powerful capabilities, real-world implementation requires rigorous engineering to ensure accuracy and compliance.

The core challenge lies in transitioning from theoretical AI potential to production-grade reliability. Kulkarni emphasized that traditional large language models (LLMs) are insufficient for tasks requiring high-stakes decision-making. Instead, organizations must build multi-agent systems that can dynamically retrieve information, verify facts, and synthesize data into structured reports. This approach is essential for industries where errors can lead to massive financial losses or regulatory penalties.

Key Takeaways from the Presentation

• Deep research agents use dynamic reasoning to perform multi-step internet searches and data synthesis.
• Healthcare and pharma require systems that ensure transparency, reliability, and strict regulatory compliance.
• Traditional LLMs fail at complex discovery tasks without additional agent-based orchestration layers.
• Task engineering and agent loops are critical techniques for improving output quality and consistency.
• The average cost to bring a new drug to market remains extremely high at approximately $2.6 billion.
• Nearly 50% of research efforts lack prior evidence due to significant barriers in accessing existing knowledge.

The High Cost of Inefficiency in Drug Discovery

The pharmaceutical industry faces a unique set of economic pressures that make AI adoption both urgent and risky. Kulkarni pointed out that bringing a single new drug to market typically costs around $2.6 billion. This staggering figure underscores the need for efficiency gains through automation. However, current research methods are plagued by inefficiencies. Approximately half of all research initiatives proceed without adequate preliminary evidence, simply because relevant data exists in silos that are difficult to access.

Researchers often struggle to connect disparate pieces of information scattered across internal databases and public internet sources. This fragmentation leads to redundant studies and missed opportunities for breakthroughs. By leveraging deep research agents, companies can automate the initial stages of literature review and data correlation. These systems can scan millions of documents, identify patterns, and highlight potential connections that human researchers might overlook. This capability transforms the early phases of drug discovery, making them more data-driven and less prone to human bias.

Overcoming Data Silos with AI

• Internal corporate data often remains isolated from public scientific literature.
• Manual literature reviews are time-consuming and prone to human error or oversight.
• AI agents can cross-reference multiple sources simultaneously to validate findings.
• Automated synthesis reduces the time required for initial hypothesis generation.
• Compliance checks can be integrated directly into the agent's workflow.

Engineering Reliability Through Agent Loops

Deploying AI in sensitive environments requires more than just prompting a model. Kulkarni discussed the importance of agent loops and task engineering in creating robust systems. Unlike standard chatbots that generate a single response, deep research agents operate in cycles. They plan, execute searches, evaluate results, and refine their queries based on intermediate findings. This iterative process allows the system to correct its own mistakes and dive deeper into specific topics as needed.

Task engineering involves designing precise instructions and constraints for each agent within the system. This ensures that the AI stays focused on the objective and adheres to specific formatting or sourcing requirements. For instance, an agent tasked with reviewing clinical trial data must prioritize peer-reviewed sources and exclude anecdotal evidence. By breaking down complex research goals into smaller, manageable tasks, developers can create a modular system that is easier to debug and maintain. This modularity also enhances transparency, allowing human supervisors to trace the logic behind every conclusion drawn by the AI.

Techniques for Robust AI Deployment

1. Implement self-correction mechanisms where agents critique their own outputs.
2. Use structured output formats to facilitate easy integration with existing databases.
3. Define clear boundaries for internet access to prevent hallucination or irrelevant data retrieval.
4. Incorporate human-in-the-loop checkpoints for critical decision points.
5. Monitor agent performance metrics continuously to detect drift or degradation.

Industry Context and Future Implications

The shift toward agentic workflows marks a pivotal moment in enterprise AI adoption. While Western tech giants like OpenAI and Google provide the foundational models, the real value lies in how enterprises customize and deploy these tools. Kulkarni’s insights reflect a broader industry trend where reliability trumps raw intelligence. Companies are realizing that a slightly less capable but highly reliable system is far more valuable in regulated industries than a powerful but unpredictable one.

This approach has implications beyond healthcare. Financial services, legal research, and supply chain management are also beginning to adopt similar multi-agent architectures. The ability to conduct thorough, auditable research automatically will become a competitive advantage. As these technologies mature, we can expect to see standardized frameworks for building and auditing deep research agents. These standards will likely focus on explainability, ensuring that every piece of generated content can be traced back to its source.

What This Means for Developers and Businesses

For software engineers and product managers, the message is clear: stop treating LLMs as magic boxes. Start viewing them as components within a larger engineered system. Success depends on the surrounding infrastructure—data pipelines, validation layers, and monitoring tools. Businesses must invest in task engineering talent who understand both AI capabilities and domain-specific constraints. Ignoring these engineering fundamentals will result in fragile systems that fail under production pressure.

Furthermore, organizations should prioritize pilot projects in low-risk areas before scaling to critical operations. This allows teams to refine their agent loops and task definitions without exposing the company to significant liability. By focusing on incremental improvements and measurable outcomes, businesses can build confidence in AI-driven research processes. This cautious yet ambitious approach aligns with the high stakes of industries like pharmaceuticals, where trust is paramount.

Looking Ahead

The next phase of AI development will likely focus on autonomous verification. Current systems still rely heavily on human oversight to confirm the accuracy of generated reports. Future iterations may include built-in fact-checking modules that cross-reference claims against trusted databases in real-time. Additionally, we may see the emergence of specialized models trained exclusively on scientific and medical literature, further enhancing the precision of deep research agents.

As these tools evolve, the role of human researchers will shift from data gathering to strategic interpretation. AI will handle the heavy lifting of information retrieval and synthesis, freeing experts to focus on high-level analysis and creative problem-solving. This collaboration between human intuition and machine efficiency promises to accelerate innovation across various scientific fields. The key to success will lie in maintaining rigorous standards for transparency and accountability throughout this transition.

Gogo's Take

• 🔥 Why This Matters: The integration of deep research agents into pharma isn't just about speed; it's about unlocking billions in value by reducing the failure rate of early-stage research. If AI can cut the $2.6 billion drug development cost by even 10%, the impact is monumental for global health and corporate margins.
• ⚠️ Limitations & Risks: Hallucinations in medical contexts are not just bugs; they are liabilities. Without rigorous 'human-in-the-loop' verification and auditable trails, companies risk severe regulatory backlash. Over-reliance on automated synthesis without understanding the underlying logic can lead to catastrophic errors in clinical decisions.
• 💡 Actionable Advice: Don't deploy autonomous agents directly into production. Start by building 'copilot' systems where AI suggests research paths but humans approve the final queries. Invest heavily in task engineering now—this skill set will be more valuable than prompt engineering in the coming years.