Research

Systems research for production agents.

The questions we are working on upstream of delivery — what makes an AI agent a coherent identity, how it remembers the user it serves, how it learns from production, and how an educated agent teaches a virgin one. Four open tracks; each feeds the services we ship.

Research tracks

Synthetic Personality Skill Distillation Agent Memory Closed-Loop Knowledge Chat Orchestration Runtime AI-Native Dashboards Prompt-Native Widgets Production Agent Interfaces

In progress

Synthetic Personality

The foundation. Stable beliefs, values, and refusal patterns that define what an agent is — invariant across users, sessions, and contexts.

Prior art · Builds on Constitutional AI (Anthropic, 2022), Sleeper Agents (Anthropic, 2024), persona consistency benchmarks, and RLAIF.

In progress

Skill Distillation

The training track. An educated agent turns its tacit competence into transferable skill documents a virgin agent can run from scratch — via a mentor–apprentice loop hardened by a regression suite.

Prior art · Builds on knowledge and context distillation, skill libraries (Voyager, Eureka), apprenticeship learning, and golden-test discipline.

Draft

Agent Memory

The 1:1 relationship layer. How an agent adapts to a specific user's context, preferences, and working style — without compromising its foundation.

Prior art · Engages with MemGPT, LangMem, agent-state literature, and the broader long-horizon agent line of work.

Draft

Closed-Loop Knowledge

The growth track. How agents learn from production interactions across all users — the shared mechanism by which capability compounds over time.

Prior art · Adjacent to RLHF data pipelines, automated eval generation, and continuous-evaluation practice.

Draft

Chat Orchestration Runtime

The end-to-end architecture of modern conversational AI: model-agnostic, client-agnostic, plugin-driven runtimes that coordinate intent, context, retrieval, tools, reasoning, reflection, memory, and rendering — with the LLM as one interchangeable component.

Prior art · Builds on agent runtime, MCP execution, context engineering, RAG, OpenTelemetry tracing, and adaptive reasoning research.

Draft

AI-Native Dashboards

A study on conversational, adaptive, living dashboard interfaces — workspaces that begin as a blank canvas with a single conversational input and build themselves in real time as the user asks.

Prior art · Engages with generative UI, conversational interfaces, spatial workspaces, and persistent memory systems.

Draft

Prompt-Native Widgets

Generative, context-aware dashboard components whose logic and rendering are defined by natural-language prompts rather than hardcoded configurations — analytical surfaces that retrieve, reason, link, and adapt.

Prior art · Builds on generative UI components, semantic linking, adaptive rendering, and intelligent caching policies.

Draft

Production Agent Interfaces

The chat surface as an operating console — knowledge bases plugged in, tools connected, agents on a roster, with real-time visibility into context budget, token spend, model choice, and concrete savings opportunities.

Prior art · Builds on AI-SDK Elements, MCP tool registries, agent observability, and model-routing research.