Skip to content

Knowledge base

Verification literacy

What an engineer needs to review generated code well, and why that code needs a different kind of review, for anyone who checks agent output.

Written by

Updated:

Getting Started 2

Our position. We are for test-driven development with agents.

A model’s output stays a standing risk. No amount of context drives the chance of an invented detail to zero, so every use weighs impact, probability and your odds of catching the error.

What does verification literacy require?

  • Ask at what change size you would deploy code you had not read, and at minimum stay on top of the test code.
  • Expect invented details next to well-known concepts: an internal library, a just-released API version.
  • Supply the security feedback loop yourself, since functional failures announce themselves and security holes do not.

Why does generated code need verification?

One practitioner essay makes the case from the code’s side. LLMs are inferrers, not compilers, so their output never becomes the settled, ignorable layer that assembly did; every use of it is a standing risk assessment weighing the impact of a failure, its probability given your setup and context, and your odds of catching it before it ships (I still care about the code (opens in a new tab), 9 July 2025). One question grounds the judgment: if you were on call for this system, at what change size would you be comfortable deploying code you had not read? For most engineers the honest answer forces a floor, and the minimum worth staying on top of is the test code, in the essay’s view.

Knowing where the model is most likely to invent is part of that judgment: hallucination clusters in the sweet spot of being adjacent to a well-known concept while asking for very specific details, which is exactly what real work looks like, an internal library or a just-released API version the model was never trained on, and exactly what the reassuring one-shot demos are not (AI Native Dev - A Developer’s Guide to Not Getting Replaced by Agents (opens in a new tab), 10 March 2026). Because hallucination is the core behavior of an inferrer rather than an occasional bug, no amount of context or model power drives the failure probability to zero, which reframes the whole job: programming shifts from a challenge of scale, writing enough code, to a challenge of trust, validating enough of what the machine wrote (The Practitioner’s Guide to Agentic Software Engineering (opens in a new tab)). The bottleneck moves accordingly.

As AI writes a growing share of code, GitHub has reported 46% of new code AI-generated, with industry voices projecting a march toward 90%, the constraint shifts from writing code to the outer loop of reviewing, testing, and merging it, which makes review the load-bearing human skill rather than an afterthought (WeAreDevelopers – Evaluating AI models for code comprehension (opens in a new tab), 6 June 2025, vendor). Anthropic’s Founder’s Playbook warns that agentic tools generate code that works, not code that is secure, and that functional failures announce themselves while security holes stay invisible until exploited, so the human has to supply the feedback loop the tool lacks (Anthropic – The Founder’s Playbook: Building an AI-Native Startup (opens in a new tab), vendor).

Resources

Articles

Talks and videos