predictor · Domain-agnostic trend-detection pipelinepredictor.explorer · Cytoscape-based living map of a field — provenance on every edgepredictor.ingest-pipeline · Daily curated ingestion + LLM extraction behind quality gates, batch-AP…predictor.daily-timers · Three domains read themselves every morning — film 06predictor.two-lens-model · Landscape vs Moverspredictor.quality-gates · Zero-token CPU checks on every LLM outputpredictor.shadow-mode · A cheaper understudy model silently re-runs primary-model jobs; disagre…predictor.calibration-reports · Rolling 7-day gate pass-rate + quality-drift reporting; thresholds tune…predictor.epoch-boundary · Scores never compare across a restartpredictor.domain-agnostic-boundary · Framework/domain separation enforced by a grep audit test — domains are…predictor.domain-template · New domain onboarding from scaffoldpredictor.chatter-source-policy · Ingest-only mode for short-form sources — count mentions, skip extracti…predictor.dev-prod-split · Two clones, one data planepredictor.playwright-harness · Self-contained web testspredictor.ontology-page · Specialist-facing taxonomy reference generated from domain profilesagent-platform · Multi-agent platformagent-platform.director · Director agent — coordinates the crew, human approval in the loop, per-…agent-platform.director.telegram · Telegram-native command surfaceagent-platform.director.morning-brief · An ambient 07agent-platform.sysadmin · Reads the server's own health every morning at 06agent-platform.scout · Wanders the box unprompted at 05agent-platform.scout.session-ingest · Twelve thousand turns of agent session log, stripped to query/progress/…agent-platform.scout.leads-ledger · An append-only story queue the Scout may add to but never edit — 324 le…agent-platform.writer · Drafts field notes from a claimed lead, in a voice loaded from bottled…agent-platform.wire-editor · Triage desk for inbound wire items, with a chief's shadow pass scoring…agent-platform.ask-assistant · One assistant per solution page, each given the same facts that built t…agent-platform.blog-pipeline · Content + marketer agents drafting and shipping posts to blog.uzelhub.c…agent-platform.decision-trace · Every agent writes structured what-ran/what-it-decided/what-it-cost row…agent-platform.health-monitor · The platform checks its own pulse hourly and pages a human when it stopsagent-platform.ops-calendar · Ops calendar as code — .ics in the repo carrying drills, kickoffs, pric…agent-platform.ops-desk · Operator surfaces behind one capability URLagent-platform.n8n-stack · n8n as the automation substrate — workflows, webhook surface via ngrok,…ops.backup-system · Layered backupsops.notify-telegram · One shared Telegram channel for all agents — [AGENT]-prefixed, outbound…ops.safe-reboot · Reboot as orchestrationops.truth-map-practice · _host/README.md as cross-project source of truth with dated corrections…ops.security-architecture · The box's exposure written down and then audited against itself on a da…ops.docker-hygiene · Disk-bounded Dockerweb.studio · Editorial site — 'a manifesto disguised as a portfolio'; literary regis…web.marketing-generator · Decision-free page productionweb.field-notes · True stories from operating the systems, drafted from the lead queue, r…web.apex-sandbox · A throwaway rebuild of the whole apex, every page in reading order with…web.agents-page · The crew introduced in plain language, each with what it does, a receip…web.sister-site-doctrine · Two sites, same truths, different attention budgets — parity without ap…web.translation-patterns · Editorial→marketing translation as a studied craftweb.coverage-tripwire · Dropped-ideas tripwireweb.promotion-inventory · The box documents itselfweb.register-gradient · Five-step register gradientweb.voice-ledger · The site's voice recorded as dictated rather than inferred — one entry…web.publishing-drip · One or two URLs a week, real dates never backdated, one promotion momen…web.cycle-canvas · Hand-tuned canvas animation on the editorial home (orbit glow, delivere…method.4cs · Converse, Constrain, Construct, Curate — the working method behind ever…method.human-ai-collaboration · 100% human-AI collaboration, zero manual coding/tests, zero hardcoded d…method.dated-amendments · Docs never silently rewritten — corrections carry dates and keep the wr…
The agent said its budget was gone. It had spent 2%.
An agent will answer a question about its own machinery confidently, and get it wrong, because answering is not the same as checking.
- The morning Telegram briefing reported a spent budget after $0.28 of a $15 cap — what stopped it was an eight-iteration limit, not the money
- All the agent got was ‘budget or step limit reached’ — never which of the two fired
- A calendar crossed the 40,000-byte read ceiling on Aug 3; the grinding started Aug 6
- It found the right answer on tool call twelve. The eighth iteration was its last.
What happened
‘Budget or step limit reached’ was the whole message: two causes, no way to tell which. So it guessed. No bug and no bad data — we asked a question we already knew the answer to, then believed the answer.
The same shape, twice more
A 51,100-byte calendar met a 40,000-byte ceiling and the clip went unmentioned. A solution found on tool call twelve died with the turn, six mornings running.
What shipped
Every fix sits outside the agent: name the limit that fired, check file size from the trace, give it somewhere to write.
Want this kind of engineering on your systems?
Get in touch