Explore agents through their source code: four engineering paradigms, protocol layers, memory, self-healing sandboxes, and practical deployments. Chinese originals are published on the official WeChat account “hexclaw” (ID: hexclaw).
Explore isolated solving and verification, step-level audits, execution receipts, diverse sampling and tolerance checks, with their limitations.
From a lesson report in DingTalk to tape diagrams and Socratic questions at the desk. A manually reviewed, anonymized real example explains three solution methods, questions for parents, and the limits of separate tutoring for each child.
From Nomic’s 768 dimensions to Qwen’s 4096: manage execution contracts, budgets, dual index revisions, resumable rebuilds, and CAS cutover separately. The old index serves valid unchanged content while its replacement builds in the background.
You can change an agent's personality, but not its operating discipline. SOUL × operating manual × Skills × tool schemas: write protocols, not incantations. The four-paradigm series concludes by tracing how a system prompt is assembled layer by layer, with real source code and constants showing which constraints cannot rely on prompting alone.
A harness is literally the equipment put on working animals: permission gates, independent judges, budgets, and controlled failure. The same model with a different harness has different system capabilities. Moving from code snippets to real projects requires a complete harness, documented here with actual source and constants.
A retrieval error is one mistake in a conversation, but compound interest in a loop. Five of this pipeline’s six stages subtract noise: query expansion → broad retrieval → score-weighted RRF → relevance floor → LLM reranking → output instructions. An examination of context engineering through real source code and constants.
Prompt → Context → Harness → Loop moves the control plane outward: message → session → system → time. Each layer delegates another decision. Real code shows what is complete, what is still developing, and who decides what “done” means.
A local implementation you can inspect with git diff: three tiers, three-dimensional recall, and two-phase dreaming. Follow an “I live here” memory to see how a local agent organizes, distills, and forgets information during its sleep cycles.
Compile natural language into a self-healing sandbox script and automatically decide whether the task should run as a script or as an Agent.
Preserving and controlling local model reasoning through streaming extraction, thinking controls, 90-second timeout recovery, context compression, and multi-step tool workflows.
Engineering notes from i18n to RTL: 1400+ translation keys, a global RTL bidi fallback, and backend language directives. All three layers are needed for an out-of-the-box Uyghur experience.
A personal Skill-distillation tutorial: capture a character in Markdown, configure YAML triggers, import it into HexClaw AI, and combine role-play with practical debugging.
Have you adopted a lobster 🦞 yet? Starting today, our crab 🦀 can help with the work that lobster can do. But HexClaw aims to handle more than one conversation: it aims to support ongoing work.
One person, Claude, and five months: building a six-repository full-stack open-source AI agent ecosystem in Go. A detailed look at HexClaw's layered architecture, core protocols, and communication contracts.
Lobster, horse, and crab: three paths addressing three problems. Similar feature lists, very different product directions. As on-device AI expands, where do their architectural tradeoffs lead?
One person, spare time, six core repositories, and 680,000 lines of code. The key is a development SOP that connects design, tests, and multiple Agents into a complete feedback loop.
Configure local Ollama or a cloud API from scratch, create an agent assistant, and add knowledge bases and skills. Start using your own desktop AI assistant in five minutes.
A K12 parent lesson-preparation assistant with grade-appropriate teaching Skills, execution-backed calculation evidence, isolated learning records, spaced review, durable delivery, and removable scenario assembly.
Create a tutor, send homework photos, inspect answers and evidence, manage mistakes, and review through Desktop or a bound DingTalk private chat. Guide for v0.5.0-beta.3.