Blog
A Collection of Interesting Prompts
A set of prompts for discussing decisions: challenge the idea, make the case for it, and compare the evidence.
Harness Anything: Designing Agent Systems for Specialized Domains
Task contracts, domain evidence, and execution boundaries for agent systems that can be checked, recovered, and evaluated.
Harness Engineering: Closing the Agent Execution Loop
How a harness organizes context, tools, state, acceptance, and recovery so multi-step tasks advance from actual results.
Agent Engineering: Retries, State, and Execution Control
Engineering details of retries, tool execution, session state, context management, and task acceptance, grounded in the public implementations of Codex, OpenCode, Pi, and Claude.
Long-Horizon Agents: Sustaining Progress Until the Task Is Truly Complete
From measuring long-horizon capability to the mechanisms behind action feedback, session handoff, goal control, experience memory, and reusable skills.
Progressive Disclosure for Skills: On-Demand Loading in Pi and DeepSeek Harness
How skill catalogs, instructions, and references enter model context in layers, through the implementations of Pi and DeepSeek Harness.
Prompt Engineering: Define the Task, Verify the Result
Task requirements, context, few-shot examples, CoT, ReAct, decomposition, structured outputs, and evaluation, explained through classic examples and practical boundaries.