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
Most AI agents work fine on a toy repo and fall apart on a real one.
The failure is usually not the model. It is context. Give an agent 56 tools and a large codebase and it burns its budget reading tool descriptions instead of solving anything.
limpiador is my attempt at fixing that. It is a Python CLI agent that does Git repository maintenance, the cleanup work that keeps getting pushed to next sprint until it turns into a migration.
The tools sit behind a searchable registry and load dynamically, so the agent only ever sees the few relevant to what it is doing right now. 56 tools across git, GitHub, filesystem, AST analysis with tree-sitter, and test and CI workflows, but never 56 at once.
A reviewer subagent runs on a scoped read-only registry and checks the work before anything lands. It cannot write, which is the entire point.
Context compaction with explicit eviction keeps long sessions alive instead of letting them quietly degrade.
Four tiers of tests, because an agent that edits your repository is not something you ship on vibes.
It started as a five day take-home and I have kept maintaining it since. Happy to talk through any part of the design if you are building something similar.
I kept the tone closer to a build log than an announcement, since that reads better than the usual "excited to share" opener. Say the word if you want a tighter version for X or a variant that ends with a link to the repo instead of an open invitation.
Github : https://github.com/MalikHaroonKhokhar/limpiador
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
Challenges
View all
Related posts
I built ReviewIQ because technical interviews don't always test whether you can actually review code.
ReviewIQ is an AI-powered code review interview trainer built for software engineers.
You pick a role, language, and seniority, then get a realistic PR diff with bugs intentionally planted in it.
You write your review.
Then the system grades it against the actual bugs, shows what you caught, what you missed, and gives you feedback on how a stronger reviewer would approach it.
The interesting part was building the grading system so it isn't just "AI thinks your answer is good." The bugs have a known ground truth, so the review can be evaluated against something concrete.
Built with Next.js, Supabase, PostgreSQL, OpenAI, and Lemon Squeezy.
I'm interested in building more products where AI isn't just a chat box, but an actual part of the product experience.
The known-ground-truth approach is a great product decision—it makes the feedback feel earned rather than like an opaque AI verdict. I also like that the flow tests the actual review skill instead of rewarding pattern-matching in interviews.
🤖 What if AI stopped waiting for prompts — and started working toward goals?
We’ve moved from chatbots → copilots → AI agents.
The next shift, I believe, is AI systems that can plan, collaborate, execute, and adapt on their own.
Imagine giving AI a goal instead of a list of instructions:
“Analyze this problem, research what’s needed, figure out the best approach, execute the tasks, and keep me updated.”
Behind the scenes, multiple specialized agents could:
🔹 Break the goal into tasks
🔹 Research & reason over information
🔹 Use tools and APIs
🔹 Collaborate with other agents
🔹 Validate their own work
🔹 Adapt when something fails
🔹 Deliver the final outcome
That’s the direction I’ve been exploring with WhatAIUpTo — moving from “AI that answers” → “AI that gets work done.”
The interesting question isn’t “What can AI generate?” anymore.
It’s: “What can AI accomplish without being told every step?”
We’re just getting started. 🚀
#AI #GenerativeAI #AgenticAI #AutonomousAgents #AIEngineering #FutureOfAI #MultiAgentSystems #ArtificialIntelligence
Every agent I ship now carries its own flight recorder.
BLACKBOX appends every step an agent takes to a hash-chained log. Each record's hash covers the record and the hash before it, so nothing can be edited, dropped, inserted or reordered afterwards without every later hash going wrong. The log renders to one HTML file: a 4D flight path you can scrub, and your browser recomputes every hash before it shows you the badge. There is a tamper button on the page. Press it and watch the chain break at that exact record.
The demo is a real agent under prompt-injection attack, 150 runs, 602 records, verified. Pure Python, an MCP server, and a BuildShip node with byte-identical hashes.
It is also my new flagship service: I build your agent, attack it 500 ways, and ship it with the recorder on from day one, so your customers can replay and re-verify any run.
Source, MIT: https://github.com/GreenAiSolution/blackbox
If you run agents that touch money or records, tell me what one of them does and I will tell you what a public replay of it would take.
