AI agents fail in boring, expensive ways.
They repeat the same step all night. They run up an API bill. They delete something nobody told them to touch. The public incident reports from 2026 are already piling up: a coding agent that edited 48,218 files in 103 seconds, and a production database and its backups gone in 9 seconds.
A prompt that says "be careful" does not fix this. The rule has to live in code, outside the model, where the agent cannot talk its way past it.
SteerPlane runs next to the agent and enforces limits while it runs:
from steerplane import guard
@guard(max_cost_usd=5, denied_actions=["delete_*"])
def run_agent(task):
...
If the agent goes over budget or tries a denied action, the run stops. You can also set a step limit with max_steps.
It does not filter content or judge whether an answer is true. It limits what an agent can do and spend.
pip install steerplane
Repo: https://github.com/vijaym2k6/SteerPlane
Site: https://www.steerplane.com
SteerPlane is open source. Patent Published.
I'm Vijay M, a student founder in Bengaluru. I'd like to hear which failure has hurt you most: runaway cost, loops, or bad actions?
The safety layer for the AGI Era.