In early October 2026 I expanded my X thread on evolving the coding agent Decode from a manually‑babysitted tool into an autonomous software factory. I laid out a four‑layer hierarchy, emphasized specification and verification at every step, described a memory taxonomy, a failure‑regression pipeline, and a clean‑graph construction process, and shared seven curated resources.
- Four‑layer hierarchy (Harness, Personalization, Automation, Continual Learning) – adopted to structure the transition from manual babysitting to autonomous factories.
- Results: a naïve extraction without a predefined ontology yielded 17 node types and 34 relationship types (including variations of “part_of”).
Your agentic software factory should tell you what it needs next.
Can't inspect failures? → Harness
Keep repeating context? → Personalisation
Keep starting tasks manually? → Automation
Same failures recurring? → Continual learning
A clean knowledge graph is built in 5 steps: Extract → Resolve Name → Embed → Deduplicate → Decide
Every new entity ends up in one of three places: merged with an existing node, flagged for review, or added as a new one.
- The clean‑graph pipeline is described as 5 steps.
- I listed 7 Second‑Brain resources.
- The discussion spanned Oct 7–Oct 9 2026.
- Specification + Verification are mandatory at every layer to guarantee well‑defined and validated outputs.
- Memory taxonomy comprises Internal knowledge, Context window, Short‑term memory, Long‑term memory.
- Failure handling follows a five‑step regression pipeline: record runs → cluster failures → write one evaluator per failure class → fix → replay.
- Clean knowledge‑graph pipeline follows five steps: Extract → Resolve Name → Embed → Deduplicate → Decide.