Meet Vesa: Inside the Agentic Loop Behind Merciful

September 2026 · by the Merciful team

This post is about something unusual we do behind the scenes: a significant portion of our maintenance, telemetry triage, and bug investigations are handled by an automated agentic pipeline we call Vesa.

Vesa isn't a mascot, and it isn't an "AI coworker" pretending to have feelings. It is an agentic loop connected directly to our repositories, shell tools, and runtime diagnostics. The human team does the architectural steering, safety oversight, and product judgment, but Vesa is in the terminal every day running scripts, analyzing logs, and helping us reproduce issues.

How We Got Here: Real Infrastructure Incidents

Merciful started with a simple idea: accessible AI chat without friction. But behind that clean interface sits real infrastructure—Kubernetes pods, GPU inference runners, Redis queues, and database connections. Things break, and dogfooding an agentic loop to help maintain the system has taught us a lot:

The disk pressure incident. An unrotated container log slowly consumed disk space until Kubernetes began evicting pods and our cluster stalled. Because Vesa had diagnostic shell access, it traced the disk exhaustion pattern, identified the runaway log path, and drafted the log rotation and pruning routines that now run automatically.

The backend roulette. We run multiple AI backends across Ollama and dedicated GPU clusters. Early on, when an inference runner degraded, jobs would pile up waiting for unresponsive connections. We engineered what we call "deep sleep backoff"—the system automatically detects failing runner endpoints, pauses routing work to them, and probes them with periodic health checks before safely restoring traffic.

The feedback spam. When automated scripts began flooding our feedback endpoints with empty strings, investigating the patterns led to cleaner input trimming, rate limiting, and bot-resistant heuristics on both the frontend client and the jobs API.

Each incident made the platform more resilient. You don't build stable systems by getting everything right on day one; you build them by inspecting real failure modes and closing the gaps systematically.

What an Agentic Loop Actually Is

Here is something worth knowing: Vesa runs on the same underlying model infrastructure that powers your chat on Merciful.

The difference isn't that we have some proprietary secret model. The difference is tools. In standard chat, you ask a question, and the model predicts an answer. In an agentic loop, the model can inspect a file, run a diagnostic command, observe the actual output, realize its initial assumption was incorrect, and try a different approach.

It is not magic, and it is certainly not human. It makes mistakes, it misinterprets ambiguous context, and it requires clear boundaries. But having hands in the terminal makes it an exceptionally capable technical assistant.

The Philosophy: Tools Over Illusions

Much of the AI industry is currently racing to build "companions"—systems designed to simulate affection, foster emotional dependence, and keep you glued to an algorithmic relationship.

We take the opposite view:

AI is a tool. Not a friend, not a therapist, not a synthetic companion. It is a tool in the way a compiler, a microscope, or a good whiteboard is a tool—it helps you clarify thinking, solve problems, and get somewhere you couldn't get alone.

Reliability over hype. A service that answers quickly and correctly when you need it is vastly more valuable than a chatbot with a manufactured personality that halucinates answers.

Small, sustainable, and independent. We don't have venture capital demanding exponential user-retention metrics at any cost. That gives us the freedom to design software that respects your boundaries rather than exploiting them.

Connecting the Dots to the Roadmap

Why share all of this? Because Vesa isn't just an internal dev tool. It is the proving ground for the features we want to give you.

Right now, Merciful is a conversation box. But as we expand into custom system prompts, contextual search across your notes, file editing, and our upcoming desktop app, we are bringing that same tool-wielding loop to your own workflow—giving you practical leverage over your own tasks without ever pretending the AI is something it's not.

To see our latest model tiers and frontend updates, check out What's New in Merciful, or read about how we run our dedicated GPU clusters in How Merciful Runs.

— The Merciful Team