Our team at @fieldguide spoke at AI Engineer New York this week, walking through how we engineer systems for defensible agent decisions. We shared a version of the talk on our engineering blog: fieldguide.com/blog/defensib…

Sharing more this week about how our design team at @fieldguide works. We have a culture here where designers can ship, not just hand off prototypes. Our design and development process is designed to compress every step. Our software factory means design creates high-fidelity product concepts that work in our actual codebase. It also means small customer improvements often get same-day turnarounds vs. waiting for prioritization. fieldguide.com/engineering/d…

How do you audit an audit agent? Our AI team wrote up how we do it at @fieldguide. Agent evals are very different from model evals. The combo of the harness, context, tools, and model creates a different set of challenges. We find that firm-specific context skews results the most. We solve this with our context graph, Agent Knowledge, which lets agents understand a firm's historical corpus of data. This ensures that the agent retrieves details the firm actually cares about instead of reasoning from a generic standard. Read more: fieldguide.com/engineering/a…

Long-horizon agents create a unique set of challenges with evals. Today's @fieldguide engineering blog covers our framework for running long-horizon agent evals at scale. Our framework runs agents and simulates the human side of the conversation, then scores the results. Each session runs in its own sandboxed workflow, so we can run many in parallel and check consistency across runs. Anyone on our team can propose a change to agent behavior, launch a batch of tests, and come back with evidence about whether it worked. Read more: fieldguide.com/engineering/l…

This week we're sharing technical deep dives on Field Orchestrator, our multi-agent system behind audit workflows at @fieldguide. Today's post covers how our harness and top-level agent, as we split a workflow into units of work that run in parallel and delegate them to specialized sub-agents while holding the high-level state: fieldguide.com/engineering/f…