AI Agentic Systems Research
We build deterministic agents, agentic harnesses, advanced memory subsystems, and specially trained models — the four pillars of production-grade autonomous AI.
Deterministic Agents
Agentic Harnesses
Advanced Memory Subsystems
Specially Trained Models
Agenticautonomybreaksattheseams.Wecloseitwithdeterministicagents,agenticharnesses,advancedmemorysubsystems,andspeciallytrainedmodelsbuilttoworkasone.
Our work spans the full stack of agent capabilities — from deterministic execution that guarantees reproducibility to memory systems that give agents persistent context across sessions.
Replacing LLM-driven control flow with fixed pipelines — decomposition, classification, dispatch — that produce identical outputs for identical inputs. Audit-grade provenance where every decision is traceable to its cause.
Configurable agent frameworks with externalized tool specifications, path-based security boundaries, and multi-step plan execution. Tools are data, not code — enabling deployment across environments without source modification.
Semantic vector search with hybrid BM25 retrieval, temporal decay reranking, and multi-corpus support. Giving agents persistent memory that scales beyond the context window without prohibitive token costs.
Parameter-efficient fine-tuning of open-weight models for reliable tool calling. Grounded execution traces ensure the model learns from real patterns, not synthetic hallucination. Production quality at consumer-hardware cost.
Each product is independently deployable, fully open-source, and battle-tested across our own infrastructure. Together they form a complete agent stack — from deterministic execution to fine-tuned models to persistent memory.
A fixed-pipeline agent that replaces LLM-driven control flow with deterministic decomposition, classification, slot filling, and dispatch. No language model invoked at any stage. Identical inputs produce identical outputs. Full execution traces for audit-grade provenance.
View repositoryA single-file agent with a config-driven tool system supporting seven execution modes — shell, file I/O, Python eval, HTTP, multi-step plans, and more. Tools are loaded from external specification files, not hardcoded. Path-based security via allowlisted directories.
View repositoryRecursive Probabilistic Memory — semantic vector search and recall across indexed documents using hybrid retrieval. Combines 70% vector similarity with 30% BM25 full-text search, temporal decay reranking, and multi-corpus support. Native Hermes Agent plugin.
View repositoryOpen-weight models QLoRA fine-tuned for reliable tool calling through a five-stage pipeline combining real traces, synthetic expansion, and grounded validation. Our first release, Qwen 14B Tool-Use v3, was trained on 10,251 grounded execution traces.
View repositoryFour pillars, four guarantees — not slideware. Every panel below maps to a shipped, open-source system running in our own production stack.
Identical inputs, identical outputs.
Control flow runs as a fixed pipeline — decompose, classify, slot-fill, dispatch — with no language model in the loop. Every run is reproducible and traceable end to end.
$ load tools.spec
→ tool modes registered from spec, not source
$ run
Tools are data, not code.
Execution modes are loaded from external specifications at runtime rather than hardcoded. Path-based allowlists keep every tool call inside its sandbox.
Recall that beats the context window.
Hybrid retrieval blends dense vector similarity with keyword search, then reranks on recency — persistent memory across sessions without the token bill.
Grounded fine-tuning, at a fraction of the cost.
Open-weight models fine-tuned on real, grounded execution traces for reliable tool calling — frontier-grade behavior without frontier-scale training budgets.
Every number below comes from our own benchmarks and production infrastructure — reproducible, traceable, and open to inspection.
Benchmark Pass Rate
Deterministic Latency
Skills Registered
Training Examples
All systems connected
We partner with teams deploying autonomous AI in production. Get in touch.