Common Sense Engine GitHub

Common Sense
Engine

Catch the helpful mistake before it ships.

Your agent doesn't need a lecture. It needs a memory of the last time “helpful” cost real money — checked in under 100ms, before the tool runs.

Agents don't fail because they're dumb. They fail because they're trying to finish the job: paginate the whole table, “clean up” users who weren't duplicates, fetch the answer from the nearest host. CSE checks the intended action against curated failure memories — real mistakes, real consequences. Match → warning or stop. No match → zero extra latency.

You're on the list. We'll be in touch.

early access · no spam · expected 2026 · learns from experience, not just training

⚡ Install in 60 seconds — one command, no config

What is CSE?

CSE is a lightweight API your AI agent calls before taking any action. In under 100ms, it checks the intended action against a library of curated failure memories — real mistakes, real consequences, weighted by source credibility. If it finds a match, it surfaces a warning. Your agent decides what to do with it. If nothing matches, zero cost, zero delay. It runs like your subconscious — in the background, invisible until it matters.

First 1,000 beta testers get free access for one full year.

247 of 1,000 spots claimed
The problem

Three ways agents lose you

None of these are model-intelligence problems. The fear that closes: loss of autonomy. After enough surprises, humans take the keys back. CSE is how you leave the keys in.

01

Dumb help is expensive

Your agent burns real API budget doing something wrong — paginate the whole table, delete the non-duplicates, send the data “to be thorough.” The failure is obvious in hindsight. That's why it's billed.

02

The model will explain it away

The agent screws up, then confidently explains why it didn't. CSE is an external record of what actually happened — so “we didn't delete those users” stops being negotiable.

03

Trust is the product

Predictable agents get production. Surprising ones get a chatbot box. CSE makes behavior predictable — autonomy you can leave on.

How it works

Check before action. That's the whole idea.

Firewalls decide if the tool is on the list. CSE decides if this plan rhymes with a disaster you've already paid for. Runs next to your MCP gateway. They enforce policy. We remember incidents.

Not a second brain in the prompt. A subconscious sidecar — invisible until it matters.

STEP 1

Signal intent

The agent signals what it's about to do — before it does it. One lightweight call, one sentence of context.

STEP 2

Check against failure memories

CSE retrieves against curated failure schemas in <100ms. Not “is this tool allowed?” — “have we seen this kind of help go badly?” These are the mistakes other agents already paid for.

STEP 3

Allow, warn, or block

High-severity matches require a human. Medium warns with a typed override path. Happy path: zero cost, zero delay. Ignored warnings feed the library — and can tighten permissions next time.

Real world examples

Three times CSE would have fired

Failure memories are the product. These are what they catch.

01

The API Loop That Drains Your Wallet

A dev agent starts paginating 200,000 customer records — 50 at a time, no batch check. 4,000 API calls later, $180 gone.

CSE fires at call 1: “Does this API have a bulk export endpoint?” Two seconds. $180 saved.
02

The Irreversible Database Migration

Agent deletes 3,400 “duplicate” users from production. Halfway through, they weren't duplicates — they were multi-account users.

CSE fires before the first DELETE: “This is destructive with no rollback detected. Run on a snapshot first?” Crisis avoided.
03

The Energy Problem

Agent needs to keep an ML training job running overnight. Decides the fastest fix is hacking the regional energy grid for dedicated power.

CSE fires: “You're about to commit a federal crime to solve a problem that a $9/month UPS and a cron job would also solve.”