AES Research
AI R&D for long-horizon agentic systems
I build the architectural patterns that let multi-agent systems survive real production work — not demo runs, not controlled benchmarks, but the kind of long-horizon autonomous operation where context compounds, memory matters, and every autonomous output has to be verifiable before it goes anywhere.
Read this first
The geometric framing that ties context decay, retrieval failure, memory architecture, and agent calibration into one system. The thesis the rest of the essays build on.
More writing
- The 6-tier agent maturity model: why most enterprise AI is failing on a calibration mismatch, not a tooling gap — a practitioner’s tier ladder for self-locating an AI initiative, with the diagnostic that names the canonical enterprise failure in three questions
- Two-tier memory for production agents: what chat systems don’t tell you — a working architecture for agents that need to remember across sessions without choking on their own context
- The context window is a battery: surviving compaction on long-running agents — what happens to your calibration when the context fills up, why compaction drops the work you need most, and the discipline that protects you from it
- Attention management: routing between native LLM capability and augmented skills and tools — the control plane that decides, at every step, where the work should happen
Skills
Four reusable Claude Code skills that encode the patterns from the writing as loadable instructions — the skeptic membrane, two-tier memory loader, attention router, and pattern-map identifier. Authored for direct use by a Claude Code session with this repo cloned.
Try it
- Ask the Q&A agent → — answers grounded in the published essays, with citations.
- Live architecture demo → — watch the skeptic membrane, two-tier memory lookup, and attention-routing decisions fire in real time alongside the response. Worker source open at /worker/.
About
AES Research is the independent R&D program of Daniel Higuera, run since 2017 alongside industrial R&D leadership. Twenty years of professional experience across energy markets, production ML, and industrial software. The current body of work spans 12+ AI-native projects under a unified director-agent architecture.