worklog

Lab

The shoulder-to-shoulder middle of the work. An under-specified problem turned into a modeled system, at whatever altitude. Most experiments fail. The instructive ones are kept.

  1. modelholds-up

    Organ, vitamin, or cancer: modeling a business into a system

    The fuzzy ask 'make this business AI-first' modeled into an operating system: three laws (organ not vitamin, organ not cancer, offer over client) that decide what to build and what to cut.

  2. note

    Validate the consumption before you build the source

    I built four scoped tool namespaces before checking how the runtime consumed them. The scope flag replaced the whole toolset instead of narrowing it. Wasted work. Prove the consumer first.

  3. modelholds-up

    A single agent holding everything timed out

    One CEO agent holding all context timed out; splitting into focused Head agents with distinct cadences made the fleet cheaper and more robust at the same time.

  4. post-mortemholds-up

    How four minutes of silence killed my AI agent

    An agent finished its work, then the orchestrator killed it before it could save. Liveness was inferred from stdout, so a long silent synthesis read as dead. A 15-line keepalive fixed it.

  5. logholds-up

    My agents had 157 tools and called none of them

    Agents swore their tools did not exist. The execution log said otherwise: a cold-spawn race, not a tool problem. Modeling the latency where it actually lived became an open-source MCP pooler.

  6. note

    The promise wins the click and loses the viewer

    Retention data from 90 days: financial-promise framing retains 4.79% of viewers; technical build framing holds 48 to 63%. The format follows the retention, not the hook.

  7. modelholds-up

    Your next move is hiding in your last win

    Rank signal sources by risk, not novelty. Winners have proven demand and format both. Derive from them first, then SEO, then competitors. Your last win is already 80% of your next move.

  8. logholds-up

    The marketing numbers were the story someone wanted me to tell

    Three wrong claims in my runtime comparison: a tool count from the ecosystem, not the tool; a star gap that misread capability; a token cost that was just a setting. The code corrected all three.

  9. specholds-up

    The tool description is the contract

    No enforceable enum at the boundary: the tool description carries the contract (canonical values, casing, defaults), and the tool returns raw source, not pre-baked analysis.

  10. note

    The document said five agents. The database said one.

    Documentation describes intent, not state. When you need to know what is running, ask the database, not the doc.

  11. specholds-up

    A voice is not a vibe. It is a contract.

    Extracting a voice from 152 posts into explicit rules, a few-shot, and an output contract: the same markdown file runs in any agent runtime.

  12. note

    The 'light' mode that cost six times more

    A LIGHT context mode promised token savings. Measurement showed it cost 6x more and broke a code path. Same session: lazy-loading skill bodies dropped 14k tokens per prompt.

  13. modelholds-up

    Stop patching your dependencies. Own the seam.

    Scattered monkey-patches have no version and nowhere to share. Own the seam: a versioned npm override survives updates and converges with upstream instead of forking noisily.

  14. post-mortemdead-end

    I trained a digital clone and it aged fifteen years

    LoRA identity training collapsed to an age when one descriptor dominated 100% of captions. The model learns whatever you hold constant. Vary everything you do not want memorized.

  15. post-mortemholds-up

    $1.85 and 832 steps, gone to a pod eviction

    A fine-tuning run on a rented A100 died at step 832 to a pod eviction. /workspace is ephemeral, nohup does not survive SSH disconnect, and pkill -9 python takes down more than your run.