Agents do real work now. They refactor code, plan features, and make decisions at two in the morning while you sleep. Then the session ends, and everything the agent learned — the context it built, the decisions it made, the reasons behind them — evaporates into a chat log nobody will ever scroll back through.
That's drift, at machine speed. The fix is the same one that works for humans: the work isn't finished until the spec knows about it.
The MCP: your spec, readable and writable by agents
Living Spec ships an MCP server, so any MCP client — Claude Code, Claude, or anything else that speaks the protocol — can work with your specs the way a teammate does:
- Read —
GetSpecs,GetPage,SearchPages,ListPages: the agent starts with the same context your team has, not a pasted fragment of it. - Write —
CreatePage,UpsertSection,UpdatePage,CreateTask: plans, decisions, and progress land as structured pages and tasks, not chat prose. - Branch —
EnsureBranch,ReconcileBranch,MergeBranch: this is the part we care about most.
Agents work on branches. Humans merge.
An agent connected to Living Spec never writes to your document directly. It calls EnsureBranch with a stable name for its session — the call is idempotent, so a long-running session can reconnect all day and keep landing work in the same place. Every page it writes, every task it files, every file it attaches goes to that branch.
Then you review. The branch shows up in Living Spec like a pull request for your spec: you see what the agent added and changed, you merge what's right, and you abandon what isn't. Renames, task updates, new pages — everything carries through the merge, and the document history shows exactly which revision came from which session.
Version control for intent. The agent gets autonomy; the document keeps its integrity.
The writeback kit
The MCP is the plumbing. The writeback kit is the habit.
It's a small skill you drop into your agent's environment that treats writing back as part of finishing: at the end of a working session, the agent summarizes what happened — what was built, what was decided and why, what's still open — and writes it to the relevant spec page, on a branch, for your review. Decisions stop living in scrollback. The "why" is findable the next time someone asks, whether that someone is a person or the next agent session picking up where this one left off.
That last part matters more than it sounds. Agents are only as good as their context, and the cheapest way to give tomorrow's session great context is to have today's session record what it learned.
We build Living Spec this way
This isn't hypothetical — the website you're reading was redesigned through this exact loop. The plan lives in a Living Spec document. An agent drafted it over the MCP on a branch; we reviewed and merged it. Design decisions, phase tasks, and progress updates all landed the same way — including the decision to write this post. When the session that built these pages ended, the spec already knew everything worth keeping.
The spec stayed true the whole way through. That's the product.
Try it
Connect your MCP client to Living Spec, point your agent at a document, and ask it to plan something — you'll see the branch appear, ready for review. If you want the writeback kit for your own agent setup, reach out and we'll get you going.