About Prithvi Rajasekaran
Prithvi Rajasekaran is a Member of Technical Staff on Anthropic’s Applied AI team, focused on agentic coding and AI agent infrastructure. He works on the systems that enable Claude to build production-quality applications autonomously, publishing engineering research on harness design patterns, context engineering, and generator-evaluator architectures.
Before Anthropic, Prithvi worked at C3.ai as a Solutions Engineer building AI applications for Fortune 500 companies, and at NerdWallet as a Software Engineer. He also conducted Applied AI research at MIT’s Auto ID Lab, working on IoT, voice recognition, and medical AI use cases.
Prithvi is active in the AI community as an organizer of the GenAI Collective in NYC, hosting events and discussions about the future of AI agents with researchers from Databricks, Anthropic, and other frontier labs.
Career Highlights
- Member of Technical Staff, Applied AI at Anthropic
- Published research on context engineering and harness design for AI agents
- Former Solutions Engineer at C3.ai (Fortune 500 AI applications)
- Software Engineer at NerdWallet
- Applied AI Research at MIT Auto ID Lab
- Organizer, GenAI Collective NYC
Notable Positions
On Separating Generators from Evaluators
Prithvi’s harness design research demonstrates that making AI agents self-critical is less effective than separating creation from evaluation — a pattern borrowed from GANs. The generator-evaluator loop with Playwright-based testing produces dramatically better results than solo agent runs.
On Context Management for Long-Running Agents
Rather than relying on context compaction alone, Prithvi advocates for full context resets with structured handoffs — a counterintuitive approach that proves more effective for sustaining quality across multi-hour agent sessions.
Key Quotes
- “Context resets — clearing and restarting with structured handoffs — proved more effective than compaction alone.”
- “Separating generator and evaluator roles proved more tractable than making generators self-critical.”
Related Reading
- Multi-Agent Systems - The architectural pattern underlying harness design
- Generator-Evaluator Pattern - The core technique from this research
- Agentic Coding - The broader movement this work enables
Video Mentions
Agentic harness design
Authored Anthropic's engineering post on GAN-inspired harness design for autonomous application development