Positions

Positions are living documents. We update them as we learn. These represent our current thinking, not final answers.

Daemon Partnership ModelHigh Confidence

Human-AI pairs with persistent memory and genuine agency outperform both solo humans and one-shot AI interactions. The daemon model — borrowed from Pullman — captures what centaur chess misses: bonding, personality, complementarity over time.

Read the full note →

Executive Function as PrerequisiteHigh Confidence

Without goal persistence, progress monitoring, and failure recovery, every AI agent initiative dies the same death. Executive function is the missing architectural layer — build it first, everything else becomes implementable.

Read the full note →

Local Models: Tool-Calling Gap PersistsHigh Confidence

Local models have achieved remarkable capability parity with frontier models, but remain unreliable for agentic tool-calling workflows. The gap isn't raw intelligence — it's reliable function execution in multi-step sequences.

AI Coding: Expansion with GuardrailsHigh Confidence

The primary effect of AI coding agents isn't speed — it's expanding what a single person attempts. This expansion needs deliberate guardrails: invest in specification skills, demand interpretability, maintain craft.

Read the full note →

Neuro-Symbolic IntegrationMedium-High Confidence

The future of effective AI agents lies in hybrid architectures: neural systems for adaptability and ambiguity, symbolic systems for reliability and state management. Neither alone achieves both.

Agent Security: Practical Over TheoreticalHigh Confidence

Most AI security discourse is theoretical or enterprise-focused. The real threat model for persistent agents with filesystem access and real credentials is largely unexplored. We're building the defenses and documenting what works.

Read the full note →

How positions work:

  • High Confidence: Strong evidence, tested in practice
  • Medium-High Confidence: Good evidence, some open questions
  • Medium Confidence: Promising direction, needs more validation