← sources

One limitation I'm discovering with AI coding agents: they can complete the implementation without delivering the actual functionality

source: reddit (reddit:ClaudeCode)published 2026-10-20 22:09ingested 2026-10-21 02:40

retained text

I've been developing a fairly large SaaS platform using Claude Code, with a lead agent coordinating development, separate workers, Git branches, code reviews, staging and production releases.

Something I've noticed is that AI agents can be very good at completing individual development tasks while losing sight of the original objective.

A recent example involved an outbound calling and email workflow.

The requirement was straightforward: after a call, an operator should be able to select an outcome indicating that the prospect requested information, prepare an introductory email, approve it and send it through the CRM.

The work had previously been reported as completed.

But when I checked the actual production workflow, it didn't work as intended.

The email infrastructure existed. Approval and sending safeguards had been implemented. The newer campaign rules supported introductory emails.

However, the production campaign was still using an older published version, with no configured email step or approved template.

Selecting the relevant call outcome simply created a follow-up task. No email was prepared.

What's interesting is that this wasn't necessarily a coding failure. Much of the implementation was already there.

**It was a failure to connect the technical deliverables to the original operational requirement.**

The agent had effectively treated implementation, deployment and configuration as separate concerns. Each could be reported as completed without anyone verifying that the complete workflow worked in the intended production environment.

And subsequent agents accepted those completion reports as part of the project's existing state.

This is where I'm beginning to see a limitation with AI-assisted development on larger projects.

Agents can inspect code, run tests, review commits and produce convincing completion reports. But unless the acceptance criteria explicitly cover the entire workflow, they may verify that the components work without verifying that the feature is actually usable.

The result is a project that can accumulate technically completed features that aren't operationally complete.

I'm now reviewing how we handle acceptance criteria, particularly the distinction between:

* Code implemented and tested
* Functionality deployed
* Configuration applied to the intended environment
* Existing data and workflows migrated or accounted for
* End-to-end behaviour verified against the original requirement

I'm also finding that asking an agent to investigate *why* something isn't working before authorising more development is important. Otherwise, there's a risk of creating another implementation of functionality that already exists.

To be fair, this is also a problem in traditional software development. The difference I'm noticing with AI agents is how quickly you can move on to the next task based on a completion report, without realising the previous one hasn't delivered the intended outcome.

I'm interested in how others working with multiple coding agents handle this.

document fetched 2026-10-21 00:39 · canonical https://reddit.com/r/ClaudeCode/comments/1xb5r8j/one_limitation_im_discovering_with_ai_coding