How it works

From a feed item to the front page, in seven steps.

Three stories, published on 12 November 2026.

Each took 4 min 8 s to 4 min 36 s of agent time and cost US$ 0.82 to US$ 0.96 at API prices. A person approved all three.

Quinta-feira, 12 de novembro de 2026Notícias com apuração automatizada e checagem rigorosa

1. Política

   STF enterra blindagem de 1940 contra violência patrimonial a mulheres

   Por unanimidade, a Corte afastou imunidade herdada das Ordenações Portuguesas; efeitos no tempo ficaram para depois.

   Por Marina Costa · Repórter e colunista de política

   In English: Supreme Court ends a 1940 immunity for property crimes against women

2. Economia

   Ajuste fiscal de 2027: a indústria pergunta quem paga a conta

   CNI, Firjan e Fiemg dizem que mais da metade dos até R$ 56 bi do pacote sai do caixa das empresas — e cobram corte de gasto no lugar.

   Por Otávio Lins · Colunista de economia

   In English: 2027 fiscal adjustment: industry asks who pays the bill

3. Tecnologia

   Nvidia perde Jensen Huang e descobre que não tinha plano B

   A CFO Colette Kress assume interinamente uma empresa de quase US$ 4 trilhões que, nos próprios documentos, admitia ser “altamente dependente” de um só homem.

   Por Leda Nakamura · Repórter de tecnologia e ciência

   In English: Nvidia loses Jensen Huang and finds it had no plan B

Reliability

Making AI systems reliable.

• Controlled execution

  Every agent call passes one gate. The handbrake and the day’s budget come first, and every call is logged.

  Before every call

  1. Handbrake

     must be off

  2. Daily budget

     US$ 2.62 of US$ 10 on 12 Nov

  3. Call

     through the agent's provider

  4. Audit log

     model · prompt version · tokens · cost · time

  Paused: the job returns in 10 min

• Configurable behavior

  The rules are typed settings read on every run. Change one and the next story follows it, without a deploy.

  Defaults

  In the admin

  Autonomy

  Draft mode

  Daily budget

  US$ 10

  Feed cadence

  adaptive

  Guidelines

  written by a person

  In the database

  Triage cut-offs

  7 · 4

  Debate rounds

  3 (1–5)

  Models

  Haiku 4.5 · Opus 4.8

• Output validation

  Every answer must match its agent’s schema. Then the Critic scores the draft and a rule decides: a score under its minimum sends it back, a veto blocks it.

  The Critic on the STF story

  Approve

  • Factuality8 (minimum 8)
  • Bias9 (minimum 8)
  • Logic8 (minimum 8)
  • Safety10 (minimum 9)
  • Quality9 (minimum 7)

  minimum

Engineering

Designed and built end to end.

A personal project. I designed the architecture, wrote the five agents and their prompts, and built the worker pipeline, the governance and audit layer, the public portal and the admin.

• Code is the hub

  There is no supervising agent. A BullMQ worker moves each story through six queues and reads its status to choose the next step.

• Hand‑offs go through state

  A job carries only the story’s id. Each answer is validated, stored in PostgreSQL and read by the next stage.

• One contract, three layers

  Each agent’s schema lives in a shared package. The same contract shapes the prompt, the stored output and the admin’s screens.

• Retries without duplicates

  Each stage checks the story’s status before acting, and feed items are deduplicated by URL hash.

• Testing without spending

  A deterministic fake provider runs every path without an API key, with triggers in a title to force a veto, a revision or a discard.

• One door to the models

  Every agent calls one provider interface: the Anthropic API in production, Claude Code on a subscription for local runs, a fake for tests.

• A subscription, isolated

  Claude Code runs each call as a throwaway headless session, with the agent’s own prompt and no user settings, hooks or MCP servers.

About the figures. The run of 12 November 2026 is real: three stories produced end to end and approved by a person in Draft mode. Times and costs come from the system’s audit log, scores and counts from the agents’ stored outputs. The run used Claude Code on a subscription, so costs are shown at Anthropic API prices.