GPT-5.3-Codex just dropped, and I wasted no time throwing it into a custom agent harness to see if it can actually handle complex supervision loops better than its predecessors.
Why I Built It
The announcement of GPT-5.3-Codex promised significantly better instruction following for long-chain tasks. Usually, when a model claims "better reasoning," it means "more verbose." I wanted to verify if it could actually maintain state and adhere to strict tool-use protocols without drifting off into hallucination land after turn 10.
Instead of testing it on a simple script, I built codex-agent-harness—a Python-based environment that simulates a terminal, manages a tool registry, and enforces a supervisor hook to catch the agent if it tries to run rm -rf / (or just hallucinates a command that doesn't exist).
The Solution
The harness is built around a few core components: a ToolRegistry that maps Python functions to schema definitions, and an Agent loop that manages the conversation history and context window.
One of the key features is the "Supervisor Hook." This isn't just a logger; it's an interceptor. Before the agent's chosen action is executed, the harness pauses, evaluates the safety of the call, and can reject it entirely.
Architecture
I wanted the tool definitions to be as lightweight as possible. I used decorators to register functions, automatically generating the JSON schema needed for the API.
- Registry Implementation
- Usage
class ToolRegistry:
def __init__(self):
self.tools = {}
def register(self, func):
"""Decorator to register a tool."""
schema = self._generate_schema(func)
self.tools[func.__name__] = {
"func": func,
"schema": schema
}
return func
def _generate_schema(self, func):
return {
"name": func.__name__,
"description": func.__doc__,
"parameters": {"type": "object", "properties": {}}
}
registry = ToolRegistry()
@registry.register
def read_file(file_path: str):
"""Reads a file from the local filesystem."""
with open(file_path, 'r') as f:
return f.read()
The Code
I've published the harness as a standalone repo. It's a great starting point if you want to test new models in a controlled, local environment without spinning up a full orchestration framework.
View Code
What I Learned
- Context Adherence is Real: GPT-5.3-Codex actually respects the system prompt's negative constraints (e.g., "Do not use
sudo") much better than 4.6, which often needed reminders.
- Structured Outputs: The model is far less prone to "syntax drift" in its JSON outputs. I didn't have to write nearly as much retry logic for malformed JSON.
- The "Lazy" Factor: Interestingly, 5.3 seems a bit too efficient. If you don't explicitly ask for verbose logs, it will just say "Done." Great for production, bad for debugging. I had to force it to be verbose in the system prompt.
References
