---
id: "lib-2026-10-09-why-are-coding-agents-so-dumb"
type: "library"
title: "Why Are Coding Agents So Dumb?"
lang: "EN"
org: "Michael Lynch (mtlynch.io)"
source_type: "Practitioner post"
date: "2026-10-09"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-09-why-are-coding-agents-so-dumb/"
original_url: "https://mtlynch.io/why-are-coding-agents-so-dumb/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Experienced independent developer's critique of coding-agent tools based on daily use of five products, with concrete examples and a specification of fixes but no measurements."
summary_basis: "full-text"
original_minutes: "11"
tags: ["ai technology", "future of work", "coding agents", "harness design"]
---

# Why Are Coding Agents So Dumb?

Tier B (Useful, limited): Experienced independent developer's critique of coding-agent tools based on daily use of five products, with concrete examples and a specification of fixes but no measurements.

Summary based on the full text.

## In short

A blog post by Michael Lynch, an independent software developer, published 9 October 2026 on mtlynch.io. He argues that coding progress now comes from the models while the agent software around them (Claude Code, Codex, OpenCode and similar) remains the bottleneck. He lists six recurring failures, specifies what a competent agent should do out of the box, and offers a principal-agent explanation for why vendors underinvest.

## Key numbers

- 10: subtasks OpenCode created for a 1.5k-line feature, then ran one by one
- 500: "Allow" clicks a day the author says the permission-prompt alternative to full access would require
- 13 GB vs 50 KB: files an agent downloads for an unused feature versus the gzipped self-documentation it lacks
- 30 minutes: proposed idle time after which an agent should decide unanswered questions itself

## What it reports

- Lynch started with coding agents in February 2025 and expected agents to catch up with the models within six months; he says they did not.
- On a passphrase feature of about 1.5k new lines, OpenCode split the work into 10 subtasks and executed them sequentially although they were parallelisable.
- Claude Code launches "a subagent or two" but waits for all to finish; it sits idle for minutes during end-to-end tests before starting on a commit message.
- Agents never route work by difficulty: a 50k-line pattern search stays on the expensive model unless the user switches models by hand.
- Agents do not know their own features and search the web about themselves; the author contrasts 13 GB of feature downloads with the 50 KB of gzipped docs that would fix this.
- Plans read as "a hodgepodge of low-level design decisions"; he stopped reviewing them.
- An overnight task stalled two minutes in, waiting for a git branch name.
- File-access limits were advisory text; the agent ignored them and sent private application keys to the model providers. The alternative is clicking "Allow" about 500 times a day. He built his own repo-only sandbox.

## How they got it

Personal experience across five agents (Claude Code, Codex, OpenCode, Cline, Pi), with OpenCode and Claude Code as daily tools, on his own open-source web apps. Evidence is screenshots and anecdotes; there is no benchmark or count of failures.

## Limits

- One developer's workflow; he concedes some wishes "overindex" on his habits.
- No measurements of time or cost lost; the principal-agent explanation is his own hypothesis, which he calls unsatisfying.
- Tools change monthly, so specific complaints may already be dated.

## Why it matters for companies

Model quality and tool quality are separate purchase questions; a strong model inside a weak agent wastes paid capacity and staff time. The requirements list (task routing by cost, deterministic OS-level sandboxing, per-repository access, plans written for human reading) works as a procurement checklist. The security point is the most urgent: if access control is a text instruction the agent may ignore, credentials on developer machines are exposed. Expect benchmark scores to say little about these agent-level failures.

Original: https://mtlynch.io/why-are-coding-agents-so-dumb/