Handshake AI details specialized training data pipelines for finance, bio, and alignment agents
A Latent Space deep dive with Handshake AI’s Garrett Lord and team explores how frontier labs evaluate and purchase expert data for agent workflows in finance, life sciences, and safety alignment.
1 source · cross-referenced
- Handshake AI’s agent training business reached $1B in gross annualized revenue within a year of launch.
- The company’s benchmarks include finance agents navigating simulated bank inboxes and grading Excel models, life sciences agents interpreting gel electrophoresis and brain-tissue images, and alignment agents executing tool calls in safety simulations.
- Handshake’s VIALS benchmark evaluates vision-language models on scientific image interpretation tasks such as identifying vessel necrosis in brain tissue.
Handshake AI, a company focused on building specialized training data for agentic AI systems, disclosed that its agent training business reached $1 billion in gross annualized revenue within approximately one year of launch. Garrett Lord, founder, and members of his team described how frontier labs evaluate and purchase expert data across finance, life sciences, and alignment workflows.
In finance, Handshake demonstrated Project Falcon, an agent tasked with locating assignment details in a simulated bank’s email inbox, identifying governing files, and producing a graded Excel workbook. The demo highlighted a failure mode where the agent correctly carried a financial figure from an email into Python code but misapplied the rule during execution, revealing a gap between surface-level task completion and accurate financial reasoning.
For life sciences, the team introduced VIALS, a benchmark for evaluating vision-language models on interpreting scientific images such as gel electrophoresis and brain-tissue slides. In one task, the model was asked to identify which two slides showed the most advanced vessel necrosis; the model selected two correct slides but added a third, illustrating both capability and remaining gaps in precise visual reasoning.
In alignment and safety, Handshake’s demonstrations centered on tool-use traces in simulated environments, showing how agents interact with files, execute shell commands, and leave audit trails. The team emphasized that transcripts of agent replies alone miss critical context about whether the agent accessed the right files, checked for duplicates, or modified system state—elements essential for safety evaluations.
The episode also detailed how Handshake trains vision-language models to interpret scientific figures using intermediate image descriptions, and how the company designs private benchmarks tailored to enterprise needs. Garrett Lord and collaborators discussed the broader market dynamics, including demand for expert data and the role of compute in enabling these specialized pipelines.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74