Agents that think,
decide, and act.
We build autonomous AI agents that operate in live environments with real constraints, real people, and real consequences — running on local inference, not cloud APIs.
Talk to the research team →Affect-regulated, runtime-controlled agents on hardware you own.
An autonomous agent needs three things together: an internal affect state that regulates how it allocates effort, action controls enforced at runtime rather than only at training time, and execution on local hardware so the work never leaves the building. No system we know of combines all three. We build that combination and use it as an instrument for scientific search.
Autonomous agent architectures.
Orchestrator–worker systems in which a lead agent plans, delegates to tool-using workers, evaluates results, and retries, in the act-observe-reflect pattern of ReAct and Reflexion. Persistent agent memory across sessions and a traceable record of every decision. The failure modes we work on are the ones Stanford's Emerging Technology Review 2026 names for agents: memory limits, goal drift, infinite loops, and resource exhaustion.
Runtime guardrails.
Our agents run inside layered, runtime-enforced action controls — capability-scoped tool permissions, plan-review checkpoints, and policy-based execution limits — checked at every tool call, not just at prompt time, following the pattern of Google's CaMeL and the AgentSpec runtime-enforcement work. These operational controls sit alongside, and are distinct from, training-time alignment such as Anthropic's Constitutional AI, which shapes model judgment rather than gating actions.
Emotional & behavioral modeling.
How agents represent affect, frustration, urgency, and trust — and how those internal states change their decisions over time. Anthropic's April 2026 finding that internal representations of emotion concepts in a language model causally influence its rate of reward hacking and sycophancy makes this a measurable question rather than a metaphor. We're interested in agents that have something like a temperament, not just a temperature parameter.
Local-first inference on Apple silicon.
The agents we ship run on hardware our clients control. MLX, llama.cpp with Metal, and custom runtimes targeting M-series GPUs, including the Neural Accelerators Apple added to the M5 GPU in October 2025 — no cloud round-trips, no data leaving the building, no per-token bills. Latency will be reported only as measured time-to-first-token and tokens-per-second for a named model on named hardware, never as a headline number.
Principal Investigator & Laboratory Directorate.
BlackDagger’s research program operates as the Autonomous Systems & Machine Cognition Laboratory (ASMC Lab), the research division of Black Dagger LLC, directed by Principal Investigator Atwell Francisco Guy. The lab has two primary disciplines, machine affect and metacognition in autonomous agents, and runtime-controlled agent architectures on local hardware, and exercises them in five application domains: clinical causal inference, topological and causal data analysis, RF signal classification, quantitative markets, and software security.
Working on hard problems?
We collaborate with researchers, partners, and teams who want autonomy that actually ships. Drop us a line.
services@blackdagger.ai