The problem: two agents, two skill libraries, one source of truth We run a two-agent workflow: Hermes for planning, orchestration, and verification; OpenCode for heavy implementation. Each agent h...
The problem: skills are metadata, and metadata drifts Last week we published how we run a two-agent workflow with 177 agent skills installed across our Hermes and OpenCode setups. The immediate fo...
The week in AI This week’s biggest story is a security wake-up call rather than a model release. INTERPOL’s African Cyberthreat Assessment 2026 reports that artificial intelligence is now involved...
The setup For the past few weeks we’ve been running a two-agent workflow: our main assistant (Hermes) plans, orchestrates, and verifies, while OpenCode (a provider-agnostic, open-source AI coding ...
Most MLOps content assumes abundant infrastructure: unlimited GPU time, reliable 5G, uninterrupted power, and a team that can spin up Kubernetes clusters on demand. But what if your deployment targ...
The benchmark blind spot A model scores 90% on MMLU. It reasons through physics problems, parses legal documents, and writes competent Python. You deploy it for a Kenyan agritech startup, and it c...
Introduction Over the past two months, we’ve built a solid foundation in AI agent engineering: from fundamentals and the ReAct pattern through memory systems, tool calling, and multi-agent orchest...
The mobile-first context Africa leapfrogged the desktop era. The internet arrived via SIM cards, not broadband cables. Smartphone penetration is now climbing past 50% across the continent, but the...
Why Fine-Tune? A foundation model like Llama 3 70B or Mistral generates fluent English, French, and Arabic. Ask it to write in Swahili or translate a sentence to Yoruba, and quality drops off a cl...
When industry meets intelligence In July 2026, Dangote Industries Limited announced plans to invest an additional $46 billion across its refining, cement, and fertiliser businesses between 2026 an...