Writing &
Reciting
Thoughts and talks on agent architecture, memory engineering, coding agent harnesses, and what it actually takes to run autonomous systems at enterprise scale.
Mnemonic Architecture
A deep exploration of memory engineering for autonomous agent systems — from foundational session design to swarm coordination to industry-wide convergence.
Convergent Engineering: How Everyone Built the Same Thing
Dozens of teams built the same agent memory architecture — independently, simultaneously. Examines why multi-tier memory with active consolidation has crossed from open research problem to engineering discipline, drawing on the 2025–2026 convergence across Anthropic, Google, Letta, Mem0, and solo engineers.
Cloud Ant Colonies
Agent swarms are real. But to understand how they think, coordinate, and grow — we need a better metaphor than "team." We need a colony. Explores how distributed agent swarms, shared memory, and a converging industry are building a genuinely new kind of distributed intelligence — drawing on swarm intelligence, stigmergy, and the principle that when an idea from mathematics is genuinely correct, nature was running it first.
Beyond the Session: Memory Engineering for Agent Teams
Autonomous agent teams are real — but to run them well, not just impressively, there's a level of memory engineering most teams haven't tackled yet. Covers intent-driven architecture, context-driven systems, and the memory substrate that separates coordinated agent teams from chaos with good intentions.
Harnesses Series
A look at the architectural choices behind coding agent harnesses — how they load context, cache prompts, and share resources across a team.
Harnesses: Eager vs. Just-in-Time
Coding agents face a fundamental architectural choice: load the entire codebase upfront (eager hydration) or search for files as needed (just-in-time). Compares eager systems like Cline against just-in-time approaches like Claude Code, and argues that just-in-time wins for prompt caching and team-level resource sharing in collaborative development environments.
Harnesses: Cache Economics
A closer look at the economics of prompt caching in coding agent harnesses — how cache hit rates, context reuse, and pricing models shape the real cost of running agents at scale.
Harnesses: Long-Running Sessions and the Tools That Get Us There
What it takes to keep a coding agent coherent across sessions that outlast a single context window, and the tooling built to get there.
Harnesses: Sandboxing and the Zero-Trust Principle
Why coding agent harnesses need to treat every execution environment as untrusted by default, and how zero-trust sandboxing principles shape agentic sandboxes.
DevGPT Coding Agents and Their Effect on the SDLC
Co-authored talk presenting the architecture and organizational impact of enterprise coding agents — how DevGPT reshaped the SDLC across J.P. Morgan Chase's Asset & Wealth Management division.
Transforming Developer Productivity: Rethinking the SDLC with AI Coding Agents
Demonstrated how AI coding agents are reshaping the software development lifecycle at enterprise scale. Presented DevGPT's architecture and real-world impact across J.P. Morgan Chase — covering cloud-native agent orchestration, secure sandbox execution, and developer productivity gains at scale.