Why are coding agents so dumb?

TL;DR Highlight

Author says agents, not models, are the bottleneck and lists the basics a coding agent should have.

Who Should Read

A developer who uses coding agents like Claude Code, Codex, or OpenCode daily and is frustrated by serial execution, model choice, and weak security.

Core Mechanics

Model and agent are different things
- The model is the brain that produces text, and the agent is the body that connects it to files and commands.
- The author says the models do the heavy lifting while the agents are the bottleneck.
Agents split tasks but run them one by one
- OpenCode split a roughly 1.5k-line passphrase feature into 10 subtasks and ran them sequentially.
- Claude Code spins up a subagent or two but waits for all of them before moving on.
Agents can't delegate to other models
- A 50k-line pattern check still runs on the expensive frontier model.
- The agent also never calls in a smarter model for hard tasks, so the user has to switch models and thinking levels by hand.
Plans are painful to read
- A Codex + GPT-6 Astra plan was a hodgepodge of low-level design decisions.
- The author started skipping plan review and went straight to implementation.

Community Reactions

Opinion that subagents solve most of this
- Commenters said to just tell the agent to use subagents, or subagents running Haiku/Sonnet.
- They said /loop fixes the stopping and sandboxes fix the security.
Experience that newer models delegate well
- One user asked the main model in OpenCode to pick subagent models and it judged difficulty and delegated much of the work to other models.
- Another said the latest models have juggled and delegated well over the past 3-4 months.
Experience that defaults just don't push parallelism
- On an office Qwen 3.6 server, OpenCode used sequential TODOs while QwenCode launched about 6 parallel subagents at once.
- The commenter had to serialize QwenCode requests into a single queue on the server.
Counterargument that self-doc lookup is intended
- A commenter said the system prompt's Academy Skill makes Claude check a docs catalog for Claude questions.
- Another said the harness can be ahead of the model, so the model doesn't know new features.

How to Apply

Tell the agent to use subagents and cheaper models
- Explicitly tell your current agent to run independent tasks in parallel subagents and give simple checks to a cheaper model.
- Paste the code example below and check whether subtasks actually ran in parallel.
Confine the agent to the repo directory
- Enforce file access with an OS-level sandbox, not a polite prompt request.
- Give access to one repo by default and open other repos read-only per session.
Give decision rules before unattended runs
- Before a long task, tell the agent to decide trivial things like branch names itself.
- Have it batch any truly necessary questions for the end.
Make plans start at a high level
- Ask for plans that start at a high level such as UI changes or the data model and work down to details.
- A commenter also suggested writing user-facing docs with the agent before the plan.

Try It

copy & run

# Paste into a coding agent like Claude Code / Codex / OpenCode (a simplified prompt version of the author's wish list)

Task: {{task description}}

Rules:
1. First split the task into subtasks and mark the ones that don't depend on each other.
2. Run independent subtasks in parallel using subagents.
3. Give simple work like code search or pattern checks to the cheapest, fastest model.
4. While waiting for tests, do things that can overlap, such as checking commit message conventions.
5. Decide trivial things like branch names yourself instead of asking. Batch any truly necessary questions for the end.

Plan rule: start at a high level (UI changes or data model) and work down to details.

Next step: check the logs to confirm the subtasks actually ran in parallel.

Terminology

HarnessThe software that wraps a model and connects it to file edits and command execution. The author compares the model to a brain and the agent to a body.

SubagentA helper agent the main agent hands part of the work to, like a lead assigning tasks to teammates.

SandboxAn isolated environment that limits files and network a program can access at the OS level. Unlike a prompt request, it is enforced.

Embarrassingly parallelWork whose pieces don't depend on each other, so you can simply run them at the same time.

Principal-agent problemA problem where decision-makers and actual users have misaligned interests, so direction drifts. The author says executives watch demos and benchmarks while developers feel everyday pain.