More Beyond Coding episodes

What Coding Agents Need From Engineers to Be Trusted thumbnail

What Coding Agents Need From Engineers to Be Trusted

Published 11 Nov 2026

Duration: 00:49:29

AI agents are transforming software development into loop engineering, challenging SaaS models and creating asymmetric cybersecurity threats that require automated defensive strategies.

Episode Description

What do coding agents need from engineers to be trusted? AI consultant Fausto Albers builds production software with agents and doesn't believe anyone...

Overview

Faustel Albers, AI consultant at Wonder Why, traces his path from sociology studies to cocktail bartending in London to restaurant ownership, where he learned to code during the Corona lockdown. He built what he believes was one of the first GenAI-powered chatbots in a restaurant, before function calling existed, which he calls "a grandiose failure" because the medium was inherently wrong for human interaction.

The conversation centers on how AI agents are reshaping software development, moving from context engineering to loop engineering as models internalize capabilities previously built into harnesses. Albers describes a layered memory system combining a wiki, a graph-based memory bank, and a local vector database, and explores how "emotion vectors" could express what matters to a developer when delegating to agents. He also examines the cybersecurity implications of AI lowering attack costs, referencing recent supermarket hacks in the Netherlands and France, and argues the SaaS paradigm may not survive the next 10 years.

What If

  • What if you built a three-layer memory system for your next solo project?

    • Move: Set up a project-specific wiki in markdown, a graph-based memory bank for concept relationships, and a local vector database in Docker for searchable context, mirroring Albers's layered approach.
    • Why Now?: Albers identifies "what decides what gets stored? You know, what decides what gets attention? And when do you retrieve it?" as open questions he has been building on "in the last half year," meaning the tooling is still maturing and early adopters can shape the workflow.
    • Expected Upside: As a possibility, separating project context from a main wiki could reduce "context pollution" and let you curate exactly what your AI agents see, improving output relevance without manual re-prompting.
  • What if you treated your AI agent's working state as a tunable parameter rather than a fixed prompt?

    • Move: Before each coding or research session, explicitly define the agent's state (curious for exploration, focused for execution) and curate the knowledge base it should draw from, so it "mimics what is important to you."
    • Why Now?: Albers frames this as an open question - "if you're loop engineering, if you're giving so much agency to something that is not you, how do you take responsibility for what they are doing?" - and notes he "doesn't have the definite answers yet," leaving room for solo operators to experiment.
    • Expected Upside: As a proposal, explicitly setting agent state per task could reduce scope creep ("branching off and branching off") and keep output aligned with your actual goal rather than a generic instruction.
  • What if you designed your software as disposable-first and hardened it only where operational data demands?

    • Move: Ship a working version quickly, then use production logs and incident data to identify which code paths are critical, as Albers describes organizations giving "feedback on the parts of the code that are crucial... based on the logs of where incidents happen in production."
    • Why Now?: Albers argues "we're shifting because we're building software that is, in some cases, inherently disposable. You build it, you deliver value, you throw it away."
    • Expected Upside: As a possibility, this approach could let you ship faster while still meeting safety requirements in high-risk domains, since you "judge it by what it does" rather than reading every line.

Takeaway

  • Build a layered memory system for each project: a markdown wiki for ontology, a graph-based memory bank for relationships, and a local vector database in Docker for search. Keep a separate main wiki to avoid "context pollution in one project."

  • Before delegating to an AI agent, explicitly define its working state (curious for research, focused for execution) and curate the knowledge base it should draw from, rather than relying on a single static prompt.

  • Judge code by what it does in production. Use incident logs and operational data to identify critical paths, as Albers describes organizations flagging "the parts of the code that are crucial... based on the logs of where incidents happen in production."

  • Design software as disposable-first: ship a working version quickly, then harden only the paths that production data flags as critical, since "we're building software that is, in some cases, inherently disposable."

  • At the start of any project, write a shared ontology - a wiki of agreed-upon terms and concepts - because Albers argues "you have to have a shared understanding of the words that you use" before building.

Recent Episodes of Beyond Coding

30 Sept 2026 Why Coding Agents Keep Making Your Codebase Worse

"Explores McDonald's World Menu Heist, AI's coding limitations (like poor long-term maintenance and refactoring), critiques unrealistic AI hype, and emphasizes engineering best practices, human oversight, and incremental productivity improvements."

19 Aug 2026 How Amazon Turns Real Failures Into Better AI Models

AI's evolution and challenges are explored through agentic systems like Amazon's Nova, model routing inefficiencies, simulated learning, enterprise tool integration, dynamic benchmarks, and the balance between specialized and generalized models.

More Beyond Coding episodes