Better Context, Less Chaos: How I Deal With LLM Nondeterminism
Reading Siddhant Khare's AI fatigue article made me question my own workflow. This is the approach I use to get less random output from LLMs.
Open field noteSoftware Engineer / Technical author
I build software.Then I write what held up.
I’m Lenin AGC, a Software Engineer in Colombia. I build web systems, open-source tools, and AI-assisted workflows. These field notes preserve the reasoning, failures, and operating decisions behind the work.
- 01Build
- 02Observe
- 03Write
Practice signal
SIGNAL / LIVE
- 01Experience
- 4 years in softwareCalculated from 02 Aug 2021
- 02Archive
- 14 published field notes
- 03Operating
- LAGC Tech / brand and product portfolio in formation
- 04Recurring signals
- Web systems / Linux / open source / AI agents
Latest transmission
Notes from the workbench
First-person notes from experiments, migrations, mistakes, and choices that changed the implementation.
OpenMemory: The Long-Term Memory Your AI Agents Actually Deserve
OpenMemory is a free, local, open-source cognitive memory engine that turns your AI agents from goldfish into assistants with real memory. Here's how I use it with MCP and OpenCode.
Open field noteOpenAI WebSocket Mode: The Secret to Ultra-Fast AI Agents
Discover how OpenAI's new WebSocket mode is transforming agentic loops by slashing latency and making multi-step reasoning 40% faster.
Open field noteConnected routes
Follow the evidence
The profile carries the point of view. The archive, repositories, and product surfaces carry the work itself.
01Field notesThe canonical archive for practical writing about software, Linux, open source, and AI systems.Read the archive02LAGC TechThe portfolio layer for the focused software products I build and operate. It is a brand/project in formation.Open the brand03GitHubPublic repositories, tooling experiments, open-source work, and traceable implementation history.Inspect the code
Open channel
Bring a real problem and enough context.
For engineering work, collaboration, or a concrete technical question, email and LinkedIn are the direct routes.