Research

Papers, standards, and technical research from our team.

Algorithm · February 2026

HBDS: How AI Agents Search Typed Documents

Hierarchical Budget-Constrained Descent Search — a four-phase algorithm that navigates typed knowledge trees under strict token budgets, and why it is impossible on flat text.

ARDS: An Agent-Ready Documentation Standard

An open specification for a canonical, write-once / generate-many context directory — with production results across 15 repositories and 54 agent definitions.

Knowledge Artifacts: Structured Files for Dual Human–AI Consumption

A taxonomy of five knowledge-artifact types, four properties that set them apart from ordinary documentation, and a case study of 8 repositories coordinated by 24 AI agents.

Fractal Agent Orchestration: Recursive Decomposition with Contract-Based Speculative Execution for Multi-Agent Software Generation

A recursive decomposition method for multi-agent software generation: dependency edges between modules are classified as contract-satisfiable or content-dependent, and contract-satisfiable branches execute speculatively in parallel. Analytical modeling of a 180-chunk system predicts a ~5× single-level speedup (Amdahl-bounded), rising to ~9× with multi-level cross-subsystem parallelism.

Remote Flow: Async-First Coordination and Time-Realistic Planning for Small Remote Software Teams

A process model for small remote software teams (3–10 engineers) built on sub-day timeboxed planning, async-first written coordination, and SPACE-aligned metrics captured automatically — with a pre-specified crossover evaluation design. A working draft; empirical data collection is planned for Q2–Q3 2026.

Does AI-Assisted Task Completion Affect Estimation Accuracy in Small Software Teams? A Crossover Study Design

A within-subjects crossover study design testing whether AI coding assistants change estimation accuracy in small remote teams. Existing trials measure AI's effect on speed (20–55% faster, sometimes 19% slower) but none on estimation. A working draft; data collection is planned for Q3 2026.