A quarter of agent assisted development: what held up
Twelve hundred commits in three months, the failure modes that repeated, and the review habits that turned out to matter more than the prompts.
AI
8 posts in this category, out of 41 total.
Tooling your agents so they stop guessing your architecture
One instructions file, a folder of skills and a generated OKF knowledge bundle as the knowledge layer for coding agents, why a generated bundle goes stale in a way that causes confident wrong answers, and the parts I would not automate.
Agentic PR review on omp: teaching a reviewer to read the repository
Replacing a single model call with a coding agent that greps the repository before judging a diff, and the output budget, parsing and state problems that only show up once the reviewer has tools.
Multi-agent systems: what we improved
Managing asyncio boundaries with CrewAI Flows, destination knowledge with local embeddings, reducing data-collection to a single turn, and bridging CrewAI's memory system to pgvector.
Multi-agent systems in production: what CrewAI handles and what you have to build yourself
What CrewAI gave us in production, what it did not, and the real lessons from running a hierarchical multi-agent trip-planning system with real users.
Fine-tuning a support model end to end: from SQL dump to LoRA adapter
A practical walkthrough of turning support chat SQL dumps into a fine-tuned LoRA adapter, including extraction, canonicalization, shard processing, and training decisions.
DSPy for production: solving LLM output consistency without prompt engineering
How DSPy helped fix score collapse in Sherpa when smaller self-hosted LLMs kept returning safe middle-range values instead of real differentiation.
Building an MCP server into Vortex: what it means to make your app tool-accessible
What it meant to add MCP tools to Vortex, the Laravel MCP package, how tools are structured, what good tool design actually requires, and what changed when an AI could operate the release workflow directly.