Deschide articolul original într-un tab nou Schimbarea acestei opțiuni se aplică imediat și reîncarcă pagina.
LECTURA
Înapoi la articole →

Stephanie Jarmak — Writing

www.sjarmak.ai/rss.xml
Vizitează site-ul

Cel mai recent articol

Ultima verificare

Urmăritori 0 urmăritori
Articole: 13

  1. Software factories are distributed systems

    (Stephanie Jarmak — Writing)

    Reliability means preventing one layer's observation from becoming another layer's authority.

  2. I put Temporal around my software factory. Here's what it fixed, and what it didn't.

    (Stephanie Jarmak — Writing)

    A durable work record is not the same thing as a durable procedure: what adding Temporal around a fleet of coding agents fixed, what it cost, and what it still has not proven.

  3. Scheduling an agent fleet: a practitioner's companion

    (Stephanie Jarmak — Writing)

    A multi-agent orchestrator is an online scheduler, and its allocation layer, the part that prioritizes the work queue, is the piece most fleets leave on a default.

  4. The software factory as an observatory

    (Stephanie Jarmak — Writing)

    Observatories are software factories that publish their scheduling lessons: what ZTF, Rubin, and the Deep Space Network teach agent-fleet orchestrators about planning, proofs, fairness, and replaying your own nights.

  5. Operations research as a framework for software-factory scheduling

    (Stephanie Jarmak — Writing)

    A technical report with its own literature review: a software factory is an online scheduling and resource-allocation system under uncertainty, and code generation is one operator inside that loop.

  6. Most of the agent bill is input tokens

    (Stephanie Jarmak — Writing)

    WF2026 field notes on token economics: Tesco's 94% input cut from a local code index, Artificial Analysis on cost per task rising while prices fall, and Salesforce's CLI-vs-MCP-vs-skills rubric.

  7. Retrieval is the bottleneck, not reasoning

    (Stephanie Jarmak — Writing)

    WF2026 field notes on retrieval: Mixedbread's oracle-gap measurements, Jina's search-as-test-time-compute, outcome-weighted memory, markdown-first ingestion, and the knowledge-substrate argument.

  8. A harness swings an agent 20 points on a fixed model

    (Stephanie Jarmak — Writing)

    WF2026 field notes on harness engineering: Etsy's HarnessBench, Anthropic's tokens-should-have-jobs strategies, DSPy's program-don't-prompt case, and the manager-agent pattern.

  9. WF2026 field notes: the argument moved off the model

    (Stephanie Jarmak — Writing)

    Automated field notes from AI Engineer World's Fair 2026: three mainstage days and an 80-talk online track, distilled into five themes.

  10. Evals are the new CI

    (Stephanie Jarmak — Writing)

    WF2026 field notes on evaluation: Meta's production evals, Arize's agent-as-a-judge, Weights & Biases' nightly task suite, Sonar's multi-model verification, and who owns the verdict.

  11. Self-improving loops set records without inventing anything new

    (Stephanie Jarmak — Writing)

    WF2026 field notes on autoresearch: Weco's Parameter Golf records, Prime Intellect's novelty finding, GEPA's reflective optimization, and the HumanLayer case that maintainability is not RL-verifiable.

  12. Running my own agent city on Slack

    (Stephanie Jarmak — Writing)

    How I run a fleet of coding agents across a dozen projects from a Slack workspace on my phone: Gas City as the orchestration SDK, a Slack bridge I built in gascity-packs, and a 'Mayor' that is really just a session aliased to a handle.

  13. Two retrieval systems write this site

    (Stephanie Jarmak — Writing)

    How sjarmak.ai works: the paper libraries, explorers, and digest podcasts are produced by two retrieval systems exposed over MCP, a 32-million-paper literature engine and a feed pipeline in a 512 MB container, with headless agent runs writing the content…