Long-form on the microagents architecture — dynamic discovery, decentralized coordination, and an operating model designed to fit the Kubernetes platforms developers and ops engineers already run.
From scheduled observation to a verified draft PR—then a repeat run with that PR still open. Explore three agents, discovered tools, versioned skills, actual prompts, Registry Viewer screenshots and captured Jaeger traces.
Explore the executions →A plan step can return a megabyte of JSON or hundreds of log streams — and that result still has to go back into the model, twice: once for the planner deciding what to do next, and once for the final answer. A tour of result compaction — the deterministic structural floor, LLM distillation, map-reduce over multi-megabyte results, and the structure-only digests that keep the planner oriented — and the one invariant that makes throwing data away safe. Grounded in the source, not just the docs.
Read article →One agent request — "What is the weather like in Tokyo?" — followed all the way down: a Jaeger distributed trace, OpenTelemetry spans with Gen-AI semantic conventions, logs correlated in Loki, and the registry-viewer's Execution DAG. The clean answer hid a tool call that failed with a 404, and the agent silently recovered. This is what it takes to see that.
Read article →A guided tour of TruvaG3's four-tier shared memory system — the academic research it came from, the five pipeline hooks that wire it into the orchestrator, how a fleet of agents stops repeating each other's work, and how a single chat agent finally remembers what a user said last week.
Read article →A guided tour of TruvaG3 — why it exists, how the architecture works in practice, what the framework ships with, and how to decide whether it fits your situation. Written for engineers and architects who already operate Kubernetes platforms.
Read article →An introduction to a microagents architecture in which specialized tools and agents dynamically register their capabilities with a shared registry, are addressed by logical names rather than hardcoded URLs, and coordinate without a central conductor — with orchestration inside each agent and decentralized coordination across the system.
Read article →