Sharing AI knowledge, practical implementation records, and more.
Google's Harness² framework enables AI agents to recursively improve their own harness — prompts, tools, memory, and skills — without modifying model weights, using a two-level recursive process that lets weaker models surpass stronger baselines on professional benchmarks.
The article analyzes the rising Forward Deployed Engineer (FDE) trend and argues that Palantir's Ontology — a three-layer digital twin of organizational operations comprising semantic, action, and governance layers — is the critical infrastructure enabling reliable AI agent deployment, contrasting it with traditional data platforms and outlining a multi-year adoption path.
Tencent Cloud's open-source TencentDB Agent Memory introduces a hierarchical L0–L3 memory system with symbolic short-term offloading to Mermaid canvases, achieving 61% token reduction on WideSearch and 33% on SWE-bench while preserving full drill-down traceability via node_id links, and integrates with Pi agents through a zero-code proxy.
This article provides a comprehensive technical analysis of the Claude Agent SDK, detailing its seven core capabilities — loop control, context management, tool systems with MCP, permission governance, hooks, subagents, and skills/plugins — and explains how they layer together to form a production-ready agent architecture with observable execution, safety guardrails, and domain knowledge integration.
This article details seven Pi Agent extensions that add MCP ecosystem access, language server integration, cost tracking, cache hit visualization, per-turn telemetry, session introspection, and tool output compression, explaining each component's capabilities, implementation highlights, and limitations for building more transparent and efficient AI agents.
This article analyzes Pi Agent's lack of built-in sandbox protection and reviews four community guardrail solutions—pi-guardrails, cc-safety-net, pi-permission-system, and containerization approaches—comparing their mechanisms, strengths, limitations, and recommended combinations for securing AI-driven code execution.
This article evaluates five memory components for Pi Agent that solve cross-session amnesia in AI programming agents, comparing their mechanisms, strengths, limitations, and ideal use cases with benchmark evidence and GitHub metrics.
A hands-on test of Claude Fable 5.1 and GPT-6 Astra at medium reasoning tier shows both models can handle complex 3D visualization, game development, and long-range coding tasks, but high token costs make daily use prohibitive for most users.
This article reviews five Pi Agent context management components—pi-lcm, pi-context-prune, pi-context-manager, billion-context-pi, and pi-topic-memory—evaluating their mechanisms, strengths, limitations, and ideal use cases for handling long conversations without information loss, cost overruns, or token explosion.
Anthropic's Dynamic Workflows for Claude Code let developers write JavaScript orchestration scripts that schedule dozens to hundreds of sub-agents in parallel, keeping intermediate results in script variables instead of context, enabling reproducible, observable, and reusable large-scale coding tasks like Bun's 750k-line Rust migration.
