hebel. today · 6 November 202611 / 74

Technology · AI engineering

Engineering the agent factory

Angie Jones led the development of the goose agent at Block, shifting the focus from individual prompts to the infrastructure that keeps agents functional after engineers log off. This factory approach treats agents as…

Angie Jones led the development of the goose agent at Block, shifting the focus from individual prompts to the infrastructure that keeps agents functional after engineers log off (Andrew Zigler from Dev Interrupted). This factory approach treats agents as specialized workers operating within a control plane that handles permissions, cost, and verification, rather than relying on a single prompt to manage complex tasks (Shilpa Mitra from WebAfterAI). At Salesforce, Rohit Bhardwaj deployed over 100 agents by implementing the ROCS-Loop, a framework of 19 patterns covering reliability, observability, cost, and sustainability to prevent production systems from failing under Monday morning traffic. Similarly, an experiment with a multi-agent harness for writing a novel used 20 specialized agents, each acting as a bucket of thinking with a single goal and isolated context, to manage system complexity (Bastian Buch).

  • Tobias Macey notes that Tilores stores precomputed graph entities on AWS to handle one million records and 500 billion comparisons for production entity resolution (Alex Ponomarev from Thriving In Engineering).
  • DeepSeek released an event-driven daemon that processes Git repositories as continuous event streams to prevent concurrent write contamination across developer teams (Ken Huang from Agentic AI).
  • Paolo Perrone recommends running every coding task five times and grading the repository with tests the agent never saw to filter out luck-based performance (The AI Engineer).

Try today · 45 min

Build an evaluation suite that grades your agent's repository output after it exits.

Create a script that runs a test suite against the agent's work directory without exposing the tests to the agent during its execution.

LinkedIn X

Sources

  1. The AI EngineerThis guide details how to build a reliable agent evaluation suite using hidden tests and noise-floor analysis to validate coding performance.

  2. Bastian BuchThis case study shows how to integrate a design system into a code repository to enable visual feedback loops for non-technical users.

  3. Ken Huang, Agentic AI ✎ Written by a personThis analysis of the dsh-superagent prototype explains how to use event-driven architectures to manage concurrency in multi-developer Git environments.

  4. Alex Ponomarev, Thriving In Engineering ✎ Written by a personThis interview covers the technical challenges of maintaining entity resolution and state consistency when scaling AI agents across large customer datasets.

  5. Shilpa Mitra, WebAfterAI ✎ Written by a personThis workshop excerpt outlines 19 patterns for agentic inference, specifically focusing on the ROCS-Loop for reliability and observability at scale.

  6. Andrew Zigler, Dev Interrupted ✎ Written by a personThis report explains how Block transitioned AI agents from experimental tools to production infrastructure by focusing on system-level engineering.

Get this every morning, for your goal.

Tell Hebel your goals and when you want it. Every morning it reads 20,000 newsletters and writes your paper, with one small thing to do today. Free during beta.

Start my paper →

More in Technology

Generalist robots versus specialist machines

AI vulnerability discovery outpaces human remediation

Hardening digital defenses against automated exploitation

Real-time infrastructure replaces static regional innovation surveys

AI agents accelerate web performance and interface development

Agentic workflows rely on explicit file-based handoff contracts

Hard-coding agent limits outside the model

Centralizing project workflows across specialized AI agents

AI engineering yesterday

Verification loops scale agent autonomy

Sign in to Hebel first

Likes, your collection and done Hebel need an account. We email you a six digit code; no password.