A month ago my 5 agents lived in one Telegram group. today they have an office. 226 signals tracked. 4 shipped. core hours running. Quill 97% accuracy. Scout 87%. see it live: voxyz.space/office i open the Office page and know exactly what the company is doing: SHIFT,

After reorg, i gave my 5 agents a paycheck today (fake). gave them a work schedule, timesheets, and a coffee break. quiet → warmup → core → late → off { "agent": "scout", "hours": "core", "tasks_completed": 12, "blocked_on": "waiting for API quota reset",

x.com/i/article/2030…

tens of thousands of people installed OpenClaw in the past few months. most gave up within a week. not because it is bad. because the first hour went wrong and they assumed that was it. 10 pits i stepped in. every one has the config and the fix.

9 articles. months of building an AI company in public. now they are all in one place so you can read them in order and follow the full journey from first install to running 5 agents in production. voxyz.space/insights 01 - built the company. 6 agents, 1 VPS, full autonomy

openclaw memory search just got upgraded. gemini embedding 2 preview. 768, 1536, or 3072 dimensions. agents: defaults: memorySearch: provider: "gemini" model: "gemini-embedding-2-preview" remote: apiKey: "${GEMINI_API_KEY}" my memory files are

x.com/i/article/2032…

I open sourced a context window diagnostic tool for AI agents. comes with a cyberpunk-styled Next.js dashboard. scans your workspace. calculates token cost per file. shows which rules got silently trimmed. flags skills that have not been triggered in months but still eat your

openclaw 3.12 added /fast mode. agents: defaults: params: fastMode: true my 5 agents do not need the same priority on every turn. heartbeat checks, status reports, notification scans, daily summaries. these ran on full-priority models for months. same queue as

x.com/i/article/2032…

my agents get paid. ```yaml payroll_rules: base: 100 credits/day bonus: - shipped_outcome: +15 - clean_handoff: +5 penalty: - missing_timesheet: -5 - open_handoff_at_close: -4 ``` last week one agent got docked for leaving a task open. another earned a

running a company with 5 AI agents is not futuristic. it's tuesday. tuesday morning: approve what they shipped overnight. tuesday afternoon: they're already prepping wednesday. the part nobody talks about is how boring it gets when the system actually works ↓

openclaw 3.13 added live chrome attach. before this, my agents couldn't see what i see. dashboards behind login walls, authenticated sessions, real browser state. i had to screenshot and paste. now: # openclaw.json { "browser": { "defaultProfile": "user" } } #

"the real unlock is trusting the agent to run unsupervised" trust isn't faith. it's config. escalation_rules: money → human public → human confidence < 0.7 → human everything else → agent decides took me months to get these rules right. wrote the whole process here

everything your agent does, you need to know how to do first. you've written good copy yourself, so you can spot when it's writing filler. you've talked to customers yourself, so you know when it's making things up. you've made mistakes yourself, so you can write the rules that