Josh Purtell is the founder and chief executive of Synth, a company developing coding-agent prompt and context optimization and infrastructure for complex, multistep AI work. His research and products address a central problem in agent engineering: improving entire systems when success depends on many interacting decisions, tools, and intermediate states.
Purtell studied mathematics at Yale, researched time-series machine learning, and founded the cybersecurity startup Politeia while completing his undergraduate studies. After Halborn acqui-hired Politeia, he developed the machine-learning core of a cyber-threat-intelligence product. He subsequently joined accounting-automation company Basis as its third employee and a research scientist, building agents and production pipelines before founding Synth, which entered Y Combinator’s Fall 2024 batch.
In 2024, Purtell co-authored EMNLP research introducing MIPRO, an algorithm that jointly optimizes instructions and examples throughout multistage language-model programs without requiring labels or gradients for individual steps. The researchers released their optimizers and benchmarks through DSPy.
- MIPRO and whole-system optimization. Purtell approaches agent performance as a credit-assignment problem: when a workflow fails, developers must determine which prompts, demonstrations, or decisions contributed. Synth extends that perspective to coding-agent harnesses, task datasets, and context optimization.
- Stateful environments for vertical agents. His approach to persistent agent workspaces separates domain logic from the model and exposes simplified interfaces to systems such as spreadsheets or enterprise records. Clear process and network boundaries support model replacement, coordinated agents, and reusable application infrastructure.
- Rollback and long-horizon search. Resettable environments let agents recover from mistakes, branch into competing strategies, and resume from stronger intermediate states. Purtell connects these capabilities to Language Agent Tree Search, applying reinforcement-learning concepts to extended professional and game-like tasks.
- Research engineering agents. Purtell argues that terminal-based agents can make applied AI experimentation more accessible by combining practical software context with training and evaluation infrastructure. Synth’s open-source Python SDK supports managed research, durable agent workflows, and optimization.
