Autonomy thresholds in long-horizon AI Employee tasks
A framework for measuring when an AI Employee crosses from supervised assistant to autonomous executor, with empirical results across coding, research, and operational workloads.
Frontier Research
Open research from Syntic on agent autonomy, evaluation, multi-agent orchestration, and the safety properties of an AI Workforce running at scale. We publish what we learn and ship what we publish directly into the platform our customers run.
Research areas
Four threads run through the lab — how AI Employees act autonomously, how we measure whether they act well, how teams of AI Agents coordinate, and how the entire Workforce stays safe under pressure.
Recent papers
Selected work from the Syntic research team.
Eval harnesses for non-deterministic Workforce systems
How we build evaluation suites that gate every deploy of the Syntic Workforce, why golden datasets out-perform synthetic ones, and what we learned shipping regression tiers in production.
Coordinating teams of Syntic AI Agents
Patterns for supervisor-worker, peer-to-peer, and market-style coordination among AI Agents, with results from real customer Workforces handling concurrent dispatches.
Safety properties of an AI Workforce under adversarial load
Red-team findings on prompt injection, lateral movement between AI Employees, and sandbox escape attempts — and the runtime guarantees Syntic ships against each.