The model
You already picked it. The 21-skill canon mirrors into every surface you run. The model keeps its identity and learns the contract; no tool replaces it.
Why ai-engineering
Not a platform. Nothing to sink.
ai-engineering works with the model you chose, the editor you already open and the harness that drives them. Quality, security and governance hold underneath — quietly, so you never notice them. The work goes faster, on fewer tokens.
When the ecosystem improves, what do you do?
The IDEs, harnesses and models around a tool get better every month. Gregor Hohpe names the two ways a tool on top can answer that. One sinks. One floats.
The sinking platform
keeps what the base already provides
- own runtime
- own flow
- own UI
every duplicate is a permanent maintenance bill
The floating platform
drops the duplicates, rides the base
- your model
- your IDE
- your harness
the floor is ai-eng — your work rides on top
When the base gains a capability, the floating platform drops its copy and adopts the base. After Gregor Hohpe, “The Magic of Platforms,” PlatformCon 2022.
How it works with you
It works with what you have.
Nothing here asks you to move. The tool joins the stack you already chose and holds its own weight inside it.
The editor
Your IDE stays yours. 8 surfaces exist in four tiers, and the row that degrades says so in writing. Enabling one regenerates its adapter. Removing it leaves the rest of your files alone.
The harness
The harness you already run keeps driving. One command turns a surface on, and the guard screens each tool call in the hot path it already had. Nothing new sits between the harness and your repo.
The ecosystem
Skills and tools from outside the canon are first-class, not competition. When the IDEs, harnesses and skills around it improve, ai-engineering adopts the improvement instead of shipping a rival. That is the floating strategy.
What it is not
Every “no” names what it protects.
Not an operating system.
It installs 5 guards, a contract format and the skill canon into the repository you already have. There is no runtime to adopt, nothing resident in your editor, and nothing new between you and your own tools.
Not a framework.
There is no lifecycle to learn and no flow to migrate to. Your build, your tests and your process stay exactly as they are. The contract gates the milestone, and everything between the gates is yours.
Not a hosted platform.
The one outbound read is an anonymous daily registry version check, cached for 24 hours and silent when offline. No code, prompt or receipt leaves the machine. The payload is a binary, and the state is versioned files you own.
Not a rigid flow.
The guards screen the run; they do not script it. A denial carries its reason verbatim, the milestone contract is one file you approve, and the agent works the way it already works.
Not another thing to maintain.
It builds on the IDEs, harnesses and skills that improve around it instead of reimplementing them, so the floor moves when the ecosystem moves. There is no wall to rebuild: ai-eng update rewrites its own files from the binary you already installed, with no network at all.
5guards deny before the damage
21skills, mirrored into every surface
8agent surfaces, tiers named
0hosted services to run
It doesn’t ask you to move. It floats on what you run.