Overview

Semantic Engineering is the method Accion Labs uses to make AI coding agents work reliably on large enterprise software. AI coding tools are fast on small, contained tasks. On a large enterprise application, AI coding tools make mistakes, because the knowledge they need is written down nowhere they can read. A coding agent needs four kinds of knowledge about the application: functional, design, architecture and code. In a typical enterprise application, the product owner holds the functional knowledge, the UX designer holds the design knowledge, the architect holds the architecture knowledge, and the engineering team holds the code knowledge. The knowledge reaches developers through documents, tickets, messages, meetings and someone's memory, and every handoff loses a little of it. The recurring cost of the knowledge and context lost in these handoffs, across roles, artifacts and tools, is called the Manual Translation Tax. AI makes writing code faster, and a team still ships no faster while it pays the tax. Semantic Engineering records the knowledge that governs the application once, in a knowledge graph: the features, screens, services and code that other parts depend on, with the relationships between them. The knowledge graph has four layers, one for each kind of knowledge: functional, design, architecture and code. Each layer is owned by the role that holds that knowledge, and that person is accountable for keeping the layer accurate. For an existing application, Breeze.AI's agents build the graph from the application itself. The agents parse the code, infer the architecture from its structure, trace the user journeys from the screens, and exercise the running application to capture the design. The custodians then review their layers, and for an application of more than two million lines, extraction typically takes two to three weeks. For a new application, the graph grows with the code: new items enter the graph as the code that builds them merges. Every item in the graph points back to its source: a code file, a ticket, a design frame or a document. Before any code is written, impact analysis checks each change against the knowledge graph, and agents cannot ignore what it finds. The specification still describes each change in full, and the knowledge graph governs how the change fits the application. Every change is checked against the knowledge graph before it is merged, and each check leaves evidence the team can audit. When the code changes, the knowledge graph is updated before the change merges, so the graph always matches the main branch. A prototype can be built by chatting with an AI tool, and a small application that its team can hold in mind works well with a written specification for each change. A large, complex or legacy application needs the knowledge graph, even when a single team works on it. For that work, Breeze.AI's agents carry each change: impact analysis before coding, then an update to the graph and a check against it before the change merges. People decide which work the agents take on, and every agent has a named owner. The same method applies to building new applications, changing existing ones, and replacing legacy systems. Legacy modernization uses its own form of the knowledge graph: a graph of the old system, a graph of the new one, and a specification that connects the two. Two platforms run the method: Breeze.AI for new and existing applications, and ASIMOV for legacy modernization. On one engagement with Breeze.AI across three products, deployments rose from 19 to 36 a month, and lead time for changes fell from 2.0 to 1.42 days. On other engagements, impact analysis replaced three to five days of investigation by a senior engineer on one brownfield application. A new user-interface workstream reused 53 percent of its design components in its first sprint, and one team saw 23 percent fewer defects on the same codebase. More than fifteen million lines of legacy code have been modernized with ASIMOV, across more than ten programs. In every case, the same four principles hold: one structured graph, agents bound by it, a named owner for each part, and a check on every change.

On the site: Agentic Software Engineering and Modernization, powered by Semantic Engineering · The Manual Translation Tax · The KG Sync Agent