Let the system do the heavy lifting
A good engineer removes repetitive work and lets the system carry it. Applied to delivery itself, that means designing, building, testing and shipping through coding agents that have been given the structure to work reliably.
Most of the work now runs through coding agents. That changes what the engineering job is: less typing, far more deciding what is true, writing it down where an agent will read it, and building the guardrails that make an agent's output reviewable.
The speed is a consequence, not the goal. A system gets built quickly because the agent is not rediscovering the codebase on every session — the architecture, the contracts and the invariants are already written down, the decisions already have their reasoning attached, and the tests and CI say plainly when something has broken. Take that structure away and the same agent produces work that has to be redone.
What the structure actually is
Knowledge the agents read before they act, and update afterwards. Living prose on architecture, contracts and invariants, with every non-trivial claim citing its source, and a check that says which pages a change has just made suspect. This is the pattern accreta extracts into a product.
Decisions written down before the code. ADRs carrying the reasoning and what would change it, so a human and an agent argue from the same source of truth rather than from whatever fits in a chat window.
Tooling authored for the purpose. Skills, subagents and hooks built for the task at hand, with one manifest generating the instruction file every agent expects and a doctor command that fails CI when they drift apart.
MCP as a delivery surface. When something has a useful core it ships as an MCP server beside its CLI, so an assistant can drive it directly instead of being told about it.
Review and test discipline that does not bend. The agent writes; the tests, the types and the review decide what survives. This is what separates the approach from generating code quickly and hoping.
Where it shows
The whole lifecycle runs this way — discovery, architecture, implementation, test and release. accreta went from first commit to a published package with all six planned phases complete in about a month. claude-account-switcher was built, tested and shipped with CI in a day. Radicas went from problem to first users inside a five-month programme, with around thirty ADRs behind it.
This repository is itself an instance: the brief decomposed into components, each open question closed by an ADR, and every decision written down with its reasons before the code that depends on it.
Examples
- accretaCompile knowledge into a wiki, do not retrieve it per query
- accretaShip an MCP server beside the CLI
- coding-agent-sync
- accreta-atlas
- claude-account-switcherBuilt, tested and shipped with CI in a single day.
Details
- Topics
- Agentic AI