01 : Home 02 : About 03 : Resume 04 : Blog
Back to Blog

Local Agentic Programming and Orchestration: First Impressions

Early findings from experimenting with local AI agents for coding workflows and how orchestration changes the development process.

I have been experimenting with running AI agents locally for programming tasks. Not through cloud APIs, but actual local models and orchestration frameworks that coordinate multiple agents to solve problems.

Here is what I found so far.

Why Local Matters

Cloud-based AI is powerful. But it has constraints: latency, cost, privacy, and dependency on internet connectivity. For day-to-day coding tasks, I wanted something that runs entirely on my machine, does not send code to external servers, and works offline.

The landscape has evolved fast. Tools like Ollama, llama.cpp, and various agent frameworks now make it feasible to run capable models on consumer hardware. The gap between local and cloud models is narrowing.

What I Learned About Orchestration

Orchestration is the part that surprised me most. Having a single LLM generate code is useful. Having multiple agents with different roles, a planner, a coder, a reviewer, a tester, is a completely different capability.

The key insight: orchestration is not about making agents smarter. It is about making the workflow more reliable. A single agent will hallucinate, get stuck, or produce low-quality output. A system where one agent plans, another implements, and a third reviews catches errors before they reach the final output.

What Does Not Work Yet

Local agents still struggle with:

The Practical Setup

For now, I use a hybrid approach: local agents for exploration, prototyping, and low-risk code generation. Cloud models for complex refactoring, debugging, and production code. The orchestration layer is the same either way.

The interesting part is watching the local models improve. Every new release closes the gap. I suspect within a year, most development workflows will default to local agents with cloud fallback for edge cases.

More findings as I continue testing.