Alexandros Karras

Most AI agents work in the demo and die in production. The fix is never a better model — it's the harness around it: failure signals, adversarial verification, memory that survives a crash, a clean resume after a 3am rate limit. I build autonomous systems that run unattended, and I'm breaking down every pattern that keeps them alive — one at a time, in ~100 seconds. Faceless. Build-in-public. Open code → github.com/alkas79 Follow if you ship agents.

@floki_web3 Because writing a real harness is harder. Prompt wrapping feels like control — it isn't. You can't prompt your way out of a bad architecture.

@alkasbuilds why do devs still insist on prompt wrapping everything ser

An agent that reads untrusted text should not also hold a shell and a network connection. Prompt injection isn't a model problem you can prompt away — it's a permissions problem. Least privilege kills the whole attack class.

@TheCryptoBlade Pydantic if Python, Zod if TS. Validate at the boundary, not inside the handler — the agent shouldn't even receive malformed output. That contract lives at the edge.

@alkasbuilds What schema validator do you recommend most

Parsing model prose with a regex is writing a bug that hasn't fired yet. Force structured output, validate at the boundary, reject on mismatch. The schema is the contract, not a suggestion.

@TheCryptoBlade Disk paths, always. tmpfs is seductive but you lose state on restart and debugging a 3am crash needs something that persists. Checkpointing to disk is the reliability layer.

@alkasbuilds Do you use tmpfs or actual disk paths

Two agents sharing a working tree don't fail loudly — they overwrite each other and both report success. Shared mutable state corrupts silently. One worktree per agent costs seconds and buys you a debuggable system.

@CaryPalmerr Durable state over the context window is the right frame — context limits are a state management problem in disguise.

Puppetmaster! Almost at 200 stars. Open source. Allows me to have Fable pilot other models for same quality, MUCH lower cost (SWEBench 47%). Universal for any harness. It is an agentic layer over durable state that reduces token costs and makes context windows irrelevant on a worker to worker basis. Beat every state of the art model without PM layer on repo length contextual understanding benchmarks by more than 2x (NL2 Repo Bench). PM has an orchestrator layer that allows you to determine a Pilot model that runs subprocess agents (no transcripts) who fully inherit tooling capabilities vs what MCPs limit. Links to all my benchmark data and published research paper are in the project readme as well! github.com/professorpalme…

@simonw The crab has charisma, the robot has... geometric energy. Mascot aside, curious whether the new UX actually changes how you reach for it.

So I guess Codex has a little robot now? (It's not as cute as the Claw'd crab)

@HarryTandy Prompts are the interface. Activations are the actual knobs. Once you've steered a model directly, prompting starts to feel like sending letters.

The AI skill for 2026 starts after the prompt stops working Open the activations. Find the feature. Steer the behavior directly Hugging Face has a 17-minute walkthrough on LLM steering without fine-tuning: 0:00 - Introduction 0:25 - Steering as neurostimulation 2:18 - Transformer architecture 4:25 - Linear representation of concepts 9:04 - Steering with transformers 13:43 - Finding steering vectors 14:36 - Using sparse autoencoders 16:28 - Conclusion This is where prompting turns into model debugging The article below shows how to build the same workflow yourself

a circuit breaker is your code learning to stop texting a service that isn't texting back. trip open, cool off, one probe before you trust it again. beats 10,000 timeouts and a self-inflicted outage.

Retry without backoff isn't resilience — it's a coordinated attack on your own dependency. The service hiccups, a thousand clients hammer it in sync, and you've upgraded an outage into a siege. Jitter is not optional.

I'm tearing down a full agent loop, one pattern at a time → github.com/alkas79/reliab…. Follow @alkasbuilds.

The boring reliability plumbing is the edge. It's the entire difference between a demo and a system that survives while you sleep.

Everyone has a frontier model now. So why do most AI agents still fall over?

a demo is n=1 with a friendly operator and a warm cache. ship after the eval, not after the applause.

Automate everything you can undo. Gate everything you can't. Most agent-safety design is just getting that mapping right — and most failures are a human approving harmless reads while irreversible deletes run unattended.

@DanKornas open-source is the right call for a productivity layer — proprietary tool lock-in is how you end up debugging a black box at 2am.

Magic - AI Productivity Tool open-source platform for building and deploying AI productivity tools

In agent systems the verifier matters more than the finder. Generating is cheap; knowing what's real is the moat. Code → github.com/alkas79/reliab…. Follow for one pattern a week.

6 of the 18 were false positives — killed before they wasted a minute of fixing things that were never broken.

I let an AI find 18 bugs. Then I built a second AI whose only job was to destroy every one of them.