Context Compaction Across Agent Harnesses: Why Summarization Is the Wrong Model
When agents forget after compaction, blame lost state over token ratios. How eight production harnesses manage context and keep long runs alive.
Blog
Thoughts on data engineering, GenAI, distributed systems, and software architecture. Views expressed here are my own and do not represent my employer. RSS
Recent writing
Beyond the Ephemeral Loop: The Architecture of Durable AI Agents
AI agents built on while loops lose state when processes crash. Durable agents require an embedded database and an operating system task scheduler.
Codemode vs. MCP: Code as Orchestration Replaces Direct Tool Calling
Exposing MCP tools to language models burns 30,000 tokens on startup and turns the model into a single-step processor. Pi, OpenCode, Cloudflare, and Goose replaced tool calls with code.
How Coding Agents Edit Files: Diffs, Snapshots, and Fast Apply
Applying diffs to disk breaks more agent workflows than reasoning errors. Here is how Aider, Claude Code, Cursor, OMP, DeepSeek Harness, OpenCode, and Morph modify code.
Oh My Pi (omp): The Maximalist Harness That Puts the IDE in the Loop
When Can Bölük forked Mario Zechner's minimalist Pi to create omp, he made a radical bet: harness defects are the real ceiling for AI coding. Here is the architecture behind that bet.
Archive
2026
- DeepSeek Harness, Cordis, and the Case for Spatiotemporal Agent Architectures
- Hermes Agent Postmortem: Why My Personal Agent Lives on a VPS
- The Deslop Stack Fallacy: Why Style Rules Hurt Thinking Quality
- Herdr and the Agent Multiplexer Gap: Why tmux Alone Isn't Enough
- Kimi K3: What a 2.8T Open Model Changes for Engineers
- How omp Runs Subagents Differently
- Pi Agent: The Minimal Harness That Became My Multi-Tool Glue
- Tau: The Open-Source Agent You Read Like a Textbook
- Taste Skill: Teaching AI Agents to Stop Building the Same Purple Landing Page
- Agentic vs Agentive: The Paper That Gets Agency Right and Accountability Wrong
- Formal Verification: The Proof Oracles Your AI Agent Doesn't Know It Needs
- Dissecting AI Agent Memory: A Practitioner's Postmortem on mem0
- Ponytail: Teaching AI Agents to Write Less Code (and Why It Works)
- Headroom: The Missing Layer Between Your Agent and Your Wallet
- Stop Prompting for Diagrams: Two Skills That Actually Work
- The GAN Pattern for Agents: Why Anthropic Stopped Letting Models Grade Their Own Work
- Generative UI: The Missing Layer That Kept AI Apps Trapped in Chatboxes
- MicroVMs: The Data Isolation Layer AI Agents Have Been Missing
- Stop Adding AI to Your Product. Start Deleting Everything It Replaces.
- Claude Code in Large Codebases: What Anthropic Got Right and What They Left Out
- The Unreasonable Effectiveness of HTML as AI Output
- Harness Engineering: Stop Blaming the Model, Fix the Environment
- DESIGN.md: The Missing File Between AI Agents and Consistent UI
- Claude Code's Source Got Leaked. Here's What Agent Builders Should Actually Learn From It.
- 15 Patterns That Keep Production AI Agents From Burning Down Prod
- Anam vs Tavus vs Hedra vs Simli: Choosing an AI Avatar Platform That Actually Looks Human
- Escaping AI Slop: Comparing Design Copilots From Skills to Figma Plugins
- Why Compound Engineering Is My New Default for Claude Code
- LiveKit vs Pipecat: One Feels Like Stripe, the Other Feels Like Airflow
- GSD vs SpecKit, BMAD, and Compound Engineering: Picking the Right AI Coding Framework
- Pros Don't Vibe, They Control: What Developers Reveal About AI Agent Use
- grep Is Not BM25: The Difference Between Pattern Matching and Retrieval
- Dense vs Sparse Retrieval: Why Your Vector-Only Search Is Missing 25% of Results
- Multimodal RAG: You Don't Need CLIP (Until You Do)
- How Claude Code Actually Works Under the Hood
- Workflow vs Agent vs Multi-Agent: The Architecture Decision Nobody Explains
- What Is an AI Agent? The Definition War That Reveals Our Industry's Confusion
- RAG for Dummies: What Nobody Tells You About Building Search That Works
- UV: The Python Tooling That Makes pip Feel Ancient
- The Rust Rewrite: Modern Developer Tools I Lean On in 2026
- 5 Underhyped Claude Code Tips That Actually Ship Code
- Browser extensions: Plasmo Has 12k Stars. I'm Still Recommending WXT
- Stop Adding Redis: Fix Your Missing Postgres Indices First
- The Fish Shell Terminal Stack: From Slow to Fast
- How Claude Code's Interview Mode Changed My Workflow
2025
- Claude Code: Things That Made Me Switch from Cursor
- Beyond the Sci-Fi Brain: 6 Design Patterns That Power Real AI Agents
- The Hard Parts of Voice AI: Latency, Prompts, and the Pink Elephant Problem
- Debugging a 500 Internal Server Error in Kubernetes: A Detective Story
- I Built a $500 SaaS Boilerplate in 5 Days (And Made It Free)
- Building Voice AI Agents: Why Text Platforms Can't Just 'Add Voice'