CK @CryptoKSquare

phenomenoner.github.io/agent-harness-…

Queue architectures, supervision trees, audit systems, migration strategies — the stuff that used to take senior teams entire quarters — can now be specified in the morning and exist by dinner.

Months of engineering throughput just became a one-day commodity.

@berryxia 龍蝦走過的路再走一次

@CuiMao @DarioAmodei 還送遊戲！太佛了

這個與 #OpenClaw 剛開始時的heartbeat機制token黑洞有異曲同工之妙，但是OpenClaw是開源的，覺得怪排查一下可以防堵；#ClaudeCode 就只能靠信仰

CatFly @imyouhu

7 months ago

Claude Code 源码泄露事件后续越来越精彩了。 有人拿泄露的源码丢给 OpenAI 的 Codex 分析，竟然找到了 Claude Code 疯狂消耗 token 的元凶——autoCompact（自动上下文压缩）机制在失败后会无限重试，完全没有上限。据源码注释记录，曾有会话连续失败高达 3272 次。 修复方法简单到离谱：加一个

126 683 6K 1.0M 4K

In traditional software, guidelines are just documentation. In AI agent stacks, guidelines are executable code. When an agent repeatedly fails despite the underlying tools being fixed, the root cause is often stale instructions (in prompts, skill files, or memory) overriding the correct path. To fix this, engineering teams must treat guidance updates with the same rigorous incident response as code-level bug fixes. linkedin.com/pulse/ai-agent… via @LinkedIn #AIAgents #AgentOps #SoftwareEngineering #LLMOps #TechLeadership #ArtificialIntelligence #DevOps

Nothing exposes the fragility of an AI workflow quite like buying a new laptop. You migrate your files, fire up your trusted AI assistant, and suddenly realize its “memory” has completely reset—because it was quietly anchored to a specific file path, shell environment, or background process on your old machine. "Portability Isn't Optional" dives into the mundane, unglamorous plumbing failures that turn local-first AI from a reliable archive into a brittle demo. If you’ve ever lost trust in an agent because its memory forked into a parallel universe, this piece explains why "local-first" isn't a mood—it's a contract. And if you can’t export and inspect your system's memory, you don’t actually own it. #AgenticAI #AIAgent #AgentWorkflow #AIWorkflow medium.com/p/portability-…

AI tooling often feels like a superpower in the moment, only to dissolve into a scattered mess a week later when you actually need to find a past decision. While the industry obsesses over larger models and “better prompts,” the true bottleneck of AI-assisted work is much more critical: retrievability. When an agent can’t reliably recall what a team has already solved, “intelligence” quickly turns into attention debt and silent duplication. "The Archivist" explores why building a smarter agent is useless if your workflow can't remember its own conclusions—and why treating memory as hard infrastructure is the ultimate form of respect for your team's time. #AgenticAI #AIAgent #AgenticWorkflow #WorkflowAutomation medium.com/p/the-archivis…

Here's a kid who's 15 and it made over 30k$ in contracts using OpenClaw (min 41:55) And now I'm curious - what are the most amazing use cases you seen using OpenClaw?

ClawCon @clawcon

7 months ago

ClawCon Austin x.com/i/broadcasts/1…

18 53 608 606K 809

Lately, the easiest AI demo to sell isn't a smarter model—it’s a prettier org chart. We put "PM Agent," "Architect Agent," and "Security Agent" on a slide, and suddenly the workflow feels mature. But assigning an AI a job title doesn't magically create code validation, rollback paths, or exception handling. It just creates "management theater." When building real-world applications, a handoff isn't safe just because it sounds specialized. It’s safe because it has a strict contract. True multi-agent reliability doesn't come from personas; it comes from typed schemas, diffs, and explicit human oversight. It's time to stop simulating startup teams inside a prompt stack. Here is why the "Multi-Agent Org Chart" is a category error, and what we should be building instead: 👇 linkedin.com/pulse/we-build… #GenerativeAI #MultiAgentSystems #AgentOps #SystemArchitecture #BuildInPublic

In OpenClaw, every model call is a composite: system prompt + tool schemas + injected files + session transcript replay + tool outputs + headroom for the response. Two practical truths fall out of this: 1. Your steady-state burn is usually the growing session transcript. 2. Your spike risk is unbounded tool output (logs, diffs, scraped pages) that gets persisted and replayed forever. Compaction helps, but it’s not magic; it’s lossy and can become a failure mode itself. #OpenClaw linkedin.com/pulse/how-open…

7) Context management is operations work The instinct when agent output degrades is to blame the model or shop for a bigger one. But in most teams running real workloads, the bottleneck isn’t capability — it’s the accumulated weight of everything the model has to re-read before it can think. Treating context as operations work is the cheapest reliability lever most teams haven’t pulled.

6) Three habits that reduce the tax right now You don’t need a perfect architecture to reduce the context tax. Start with habits that force signal-to-noise: 1. Write the “why,” not just the “what.” One paragraph per decision beats 50 lines of debate buried in chat. 2. Index-first retrieval. Ask for file lists, diffstats, counts, or a table of contents before you pull full content. 3. Use handoffs. When a session drifts, start a fresh one with a short note: what was decided, what was tried, what happens next. The common thread is treating context as an operational concern — something you budget, clean, and hand off deliberately — not a side effect of chatting.

Most teams blame “the model” when agent work starts to feel expensive and slippery. In practice, the bill is often continuity: the growing bundle of instructions, tools, files, and history the model must re-ingest on every turn. Bigger context windows help, but they don’t remove the underlying constraint: attention is finite, and long context can hide the important parts. If you don’t govern context, you pay three ways: money, latency, and quality drift. #OpenClaw #AIAgnets #AICoding