What is an initializer agent?
An initializer agent is the first session of a long-running agent build, run once to set up the environment, a full feature list and a progress log so that later coding sessions can make incremental, verifiable progress.
Anthropic introduced the pattern in "Effective harnesses for long-running agents" (2025-11-26). The starting problem: a frontier model running in a loop with compaction still fell short of building a production-quality clone of claude.ai from a high-level prompt. Each new context window began with no memory of what came before, and the agent either tried to do too much at once or guessed at the state it inherited.
The fix splits the job. An initializer agent runs once and leaves behind a structured environment. Every later session is a coding agent that reads that structure, does one piece of work and leaves the environment clean for the next one.
What the initializer creates
A start-up script
An init.sh that starts the development server, so no later session wastes time working out how to run the app.
A progress log
A plain text file where each session records what it did, read first by the next session.
A feature list in JSON
More than 200 features for the claude.ai clone, each with a category, description, test steps and a passes field set to false. JSON was chosen because the model was less likely to rewrite it than Markdown.
A first commit
An initial git commit, so every later change has history to read and to revert to.
What each later session does
The coding agent gets its bearings by reading the working directory, the progress file, the git log and the feature list, then picks the highest-priority feature still failing. It runs init.sh and a basic end-to-end test to catch anything the last session broke, works on one feature, and may change only that feature's passes field. Removing or editing tests is ruled out in the prompt. It ends with a descriptive commit and a progress update.
In Anthropic's footnote, the initializer differs from the coding agent only in its first prompt. The system prompt, tools and harness are the same, which keeps the pattern cheap to adopt.
The failures it was built to stop
One-shotting
Building too much at once, running out of context mid-feature and leaving undocumented half-work.
Premature victory
A later session seeing progress and declaring the project finished. A full feature list set at the start removes the room for that.
Untested completion
Marking features done after unit tests or a curl check. Prompting the agent to test as a user through browser automation caught bugs the code alone did not show.
Known limits
Anthropic notes the browser tooling could not see native alert dialogs, and that it is still open whether one general coding agent or several specialised ones works better.
How Motiontic's engineers apply it
On long builds, Motiontic's engineers write the initializer's outputs by hand or review them before any coding session runs: the feature list is effectively the scope document, so a person signs it off. The progress file and start-up script are then treated as part of the codebase and kept current by every session.
Initializer agent compared with the coding agent that follows it
| Initializer agent | Coding agent | |
|---|---|---|
| Runs | Once, in the first session | Every session after that |
| Main output | Environment, feature list, progress log, first commit | One finished feature and a commit |
| Reads | The high-level prompt | Progress file, git log, feature list |
| May change | Everything, it is creating the structure | Only the passes field of the feature it worked on |
Common questions
What is the initializer agent pattern?
It is a two-part harness for long builds: one agent sets up the environment, feature list and progress log once, then a coding agent makes incremental progress each session from that shared state.
Is an initializer agent the same as a planner agent?
They overlap. Anthropic's later harness work uses a planner that expands a short prompt into a product spec. The initializer goes further and sets up the runnable environment and the tracking files.
Why use JSON for the feature list?
Anthropic found the model was less likely to change or overwrite a JSON file inappropriately than a Markdown one, which matters when the list is the record of what is done.
Related
Long-running agent
Agents that work for hours or days, and the state that lets them pick up where they stopped.
"Done when" criteria
Completion written down as checks before the agent starts, not judged after.
Sprint contract
Generator and evaluator agree on done before any code is written.
Harness engineering
Designing the environment an agent works in, so its output can be checked and shipped.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.