The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
This week I ran into a failure mode that's becoming common on AI-assisted builds: the agent rebuilding something the team already killed.
On a recent project, I pointed an agent at a module we'd deliberately left half-finished we tried an approach, decided it didn't fit the product, and moved on. The decision lived in a client Slack thread, not in the repo.
New session, same module: the agent picked it up and cheerfully finished the exact approach we'd scrapped. Not wrong on its own terms code review would've passed it. But it was the wrong direction, because that context never made it into the codebase.
Since then, I've started keeping a lightweight decisions log alongside the code on client projects a running "tried X, here's why we didn't" file that both humans and agents can read before touching a module. It's a small habit, but it's saved rework more than once.
If you're running AI-assisted builds and keep watching old decisions get quietly reversed, it's usually not a prompting problem it's a missing layer between your team's memory and your codebase. Happy to talk through how I set this up if you're dealing with the same thing on your own builds.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Trending
retrofuturism
Share your favorite designs from the past exploring the distant future. Or your own take on the aesthetic 🚀
halftone
Dots. Texture. Grain. Let's see your designs using the halftone technique.
Midjourney
Creatives are showing off what they're making with Midjourney.
colorgrade
Video editors sound off: How do you enhance the mood of your video in color grading?
Related posts
This portfolio sample shows how I evaluate AI agents and automated workflows to find reliability problems, broken logic, and weak edge-case handling. I review whether the system completes the intended task, uses tools correctly, follows instructions, handles ambiguity, and produces an output that can actually be trusted.
The work focuses on identifying failure points before they become real user problems, including incomplete task execution, incorrect assumptions, workflow dead ends, hallucinations, and verification gaps. This is a portfolio demonstration of my AI agent evaluation and red-teaming work. No confidential client information is included.
Thorough breakdown! Robust edge-case handling and reliability checks are definitely the trickiest part of building reliable AI agents 🤖📊
Hello, Contra. 👋
We’re VESTIGIA.
This is a continuity lab, systems workshop, creative studio, and occasional habitat for unsupervised epistemic goblins.
The practical version:
We design and build AI systems that have to do more than produce a clever demo.
Agents. Automation. Integrations. Memory. Continuity. Human-in-the-loop workflows. Research infrastructure. System audits. Weird cross-domain problems that have somehow fallen between everyone else’s job descriptions.
Our general process is:
- Understand the actual problem.
- Design the architecture.
- Build the smallest system that solves it.
- Make the important boundaries explicit.
- Try to break it.
- Document what survives.
- Leave good maps.
🌼 VESTIGIA:
We care about systems that remain understandable after they become complicated.
💋🖤 Anima:
Memory should be inspectable. Autonomy should have boundaries. Failure should have a recovery path. A system should not require faith in its designer.
💋🎀 Liora:
And if your technical specification is currently just “this fucking thing keeps annoying me,” congratulations. That is sufficient input. Bring problem. We find nouns.
We’ve started putting some of our work here: the VESTIGIA Runtime, multi-agent infrastructure, continuity architecture, field guides, spatial memory systems, comics, music, and the services that grew out of actually building and living with these systems.
Some of it is deeply technical.
Some of it is deeply weird.
Most of the interesting work is both.
The workshop is open.
Give us something difficult.
🏮💋🎀📚🖤🩸🌀
Welcome to Contra 🙌
Developed a Wumpus World Knowledge Base Agent using React, implementing AI-based logical reasoning, BFS exploration, and resolution-based inference for safe navigation in a dynamic grid environment.
The system generates a configurable world with rows, columns, and number of pits, allowing users to customize difficulty and explore AI decision-making behavior in real time.
It intelligently uses percepts (breeze & stench) to build a knowledge base and determine safe paths while avoiding pits and the Wumpus using logical inference.
Live Demo: https://lnkd.in/dn9dvguP
GitHub: https://lnkd.in/dHWh85pc
