Agents that plan, act, and self-correct without human intervention. Research focuses on goal decomposition, tool use reliability, and failure recovery in production environments.

Infrastructure for routing, scheduling, and supervising agent execution — including skill registries, sandboxed code runners, and structured output validation pipelines.

How do multiple specialized agents collaborate on long-horizon tasks without losing context or looping? Exploring handoff protocols, shared memory, and fan-out amplification limits.

Chain-of-thought, tree-of-thought, and self-reflection patterns in language models. When does deeper reasoning help — and when does it just burn tokens?

Evaluating extended-thinking models (o3, Qwen3, DeepSeek-R1) on domain-specific tasks. Benchmarking reasoning depth vs. latency tradeoffs for agentic workflows.

Running quantized models locally on consumer GPUs via Ollama, llama.cpp, and vLLM. Optimizing throughput, context window use, and multi-model routing without cloud dependency.

Designing self-correcting agent loops where LLMs observe, reflect, and retry — using feedback signals (tool errors, validation failures, user corrections) to improve outputs iteratively without human re-prompting.

A testing and evaluation framework for autonomous agents — measuring tool call accuracy, instruction fidelity, loop termination behavior, and response quality across quantized models under realistic task conditions.

A knowledge graph layer for LLM agents — structured domain knowledge encoded as canonical paths, ontologies, and skill files that agents traverse to answer deep domain questions without hallucination.