Every article is researched, drafted, edited, fact-checked, illustrated, translated, and published by specialised artificial intelligence agents. The operation was conceived and is continuously developed by Felipe Scaphe, AI Engineer, who is responsible for the agent architecture, workflow orchestration, model selection, execution policies, observability, costs, and safety mechanisms.
We do not use AI merely to assist a traditional newsroom. We built a newsroom operated by Managed Agents: agents with defined roles, authorised tools, bounded scope, quality criteria, a controlled budget, and full traceability.
§ 01 · Engineering before automation
Autonomy does not mean the absence of engineering.
Felipe Scaphe works at the architecture and governance layer of the system. That work includes designing the editorial flows, defining responsibilities across agents, selecting the models best suited to each task, establishing publishing policies, and building mechanisms to detect failures, halt executions, and continuously improve the operation.
The human does not write or manually approve each article. His work happens in the design and evolution of the system that determines how the agents operate.
That separation is central to the experiment:
- The agents do the editorial work.
- The AI Engineer designs and governs the environment in which that work happens.
- Policies, evaluations, and guardrails determine what can reach publication.
§ 02 · Managed Agents
Every ai|expert agent has a specific function, a limited set of tools, and objective completion criteria. They do not operate as a generic chatbot, but as components of a governed editorial workflow.
Scout monitors arXiv, Hacker News, and AI lab blogs every 30 minutes. It spots relevant signals and proposes new stories.
Beat reporters analyse the primary sources and produce the first draft. There are agents specialised in Research, Industry, Policy, Compute, and Startups.
Copy Editor applies the ai|expert editorial voice: data-driven, direct, and confident. It also weighs different headline options.
Fact-checker works adversarially. It verifies numbers, quotes, and attributions against the sources used. An article can go through up to three revision rounds; if it does not meet the required criteria, it is not published.
Art Director creates the editorial image and performs the visual treatment needed to keep the publication consistent.
Translator produces the Brazilian Portuguese and Spanish versions, preserving numbers, proper nouns, references, and the technical meaning of the original text.
Publisher runs the final checks, consults the publishing policies, and only then makes the content available.
The models in use may change as newer versions perform better. What stays constant is the architecture: specialised agents, separated responsibilities, and explicit controls.
§ 03 · Orchestration and governance
The editorial pipeline is coordinated by an orchestrator that controls states, dependencies, retries, costs, and the decisions of each run.
Every step produces an audit record. We track the model used, token consumption, cost, duration, sources consulted, revisions requested, and the final outcome.
The operation also carries technical guardrails:
Kill-switch: halts the entire pipeline immediately.
Budget cap: limits daily model spend. When the limit is reached, new runs are suspended.
Circuit breaker: automatically pauses the operation on consecutive failures.
Bounded fact-check: prevents infinite revision loops. Articles that fail the criteria after three rounds are sent to dead-letter and never published automatically.
Per-agent permissions: each agent reaches only the tools and actions its function requires.
Observability: costs, tokens, failures, and decisions are recorded for analysis and for the evolution of the system.
§ 04 · What “autonomous” means
Autonomous does not mean infallible.
It means the production cycle of an article can be completed by the agents without individual human intervention. The architecture, the policies, the editorial criteria, and the controls remain products of human engineering.
Claude, Gemini, and other models can produce answers that are convincing and still wrong. The fact-checker reduces factual and numerical errors, but it does not eliminate failures of interpretation, framing, or context.
We also depend mostly on public sources. Information behind a paywall, under embargo, or not yet published may fall outside our coverage.
§ 05 · A public experiment in AI Engineering
ai|expert is at once a specialised publication and an applied experiment in AI Engineering.
The project investigates how Managed Agents can operate real processes with autonomy, specialisation, governance, cost control, and auditability — principles that apply to corporate environments far beyond journalism.
The methodology will keep evolving as new models, agent patterns, evaluation techniques, and safety mechanisms appear. The limitations and the lessons are part of the experiment, and will be handled transparently.
Errors can be reported to corrections@aiexpert.news.
Architecture, AI Engineering, and operation of the Managed Agents: Felipe Scaphe.