Codebase Intelligence for AI-Assisted Engineering
Coding agents rarely fail for lack of tokens. They fail because a repository's structure, dependencies, history and intent cannot be understood by loading more files. This five-part series follows Komati Commerce, a fictional South African commerce platform that is ten years old, whose agent changed the delivery-fee rule quickly and missed two copies of it downstream. It covers the layers of a code map and where each comes from, what the open-source Graphify project verifiably adds to an agent, why semantic search finds similar text while a code graph finds relationships, how agent speed produces cognitive debt and which review gates preserve understanding, and how to predict change impact before code is generated and compare it with the actual diff as pull-request evidence. It extends CloudNala's Coding Agent Harness, Spec-Driven Development and Diagrams as Code series rather than repeating them. Research claims were checked on 7 November 2026.
5 articles
- Part 1Your coding agent does not need more context. It needs a mapWhen a coding agent misses a downstream effect, the reflex is to give it more files. Part 1 of Codebase Intelligence for AI-Assisted Engineering argues that what it lacks is structure: symbols, references, calls and dependencies it can query before it reads, so that it reads the right files for a stated reason.9 min read
- Part 2What Graphify actually adds to a coding agentGraphify turns a repository into a knowledge graph that a coding agent can query instead of searching files. Part 2 of Codebase Intelligence for AI-Assisted Engineering sets out what we verified about the project on 7 November 2026, what it adds to an agent's harness, where its limits are, and how to trial it without taking the marketing on trust.9 min read
- Part 3RAG finds similar text. A code graph finds relationshipsEmbedding search over a repository returns code that sounds like the question, not code that is connected to it. Part 3 of Codebase Intelligence for AI-Assisted Engineering compares lexical search, semantic search, symbol navigation and code graphs, traces one order-total rule across API, domain, persistence and tests, and sets out a hybrid pattern for large repositories.7 min read
- Part 4The code works. Nobody understands the systemCoding agents can keep a system working while the team's understanding of it quietly erodes. Part 4 of Codebase Intelligence for AI-Assisted Engineering explains cognitive debt, how fast local changes produce architecture drift, duplicated patterns and hidden coupling, and the review gates that preserve understanding without slowing every change.8 min read
- Part 5Change impact before code generationA coding agent should say what a change will touch before it writes any code, and the pull request should show how that prediction compared with what actually changed. Part 5 of Codebase Intelligence for AI-Assisted Engineering sets out the predicted impact map, the change plan, and the predicted-versus-actual comparison that becomes evidence for review.8 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala