Long-running agents that stay on course

A long-running agent — a migration, a refactor, a research sweep — drifts across sessions. It forgets the constraints you set on day one, re-proposes approaches you already rejected, and loses track of what's done versus in-progress.

This tutorial shows how to back an agent run with Jackdaw so it resumes where it left off, without drift.

What you'll need

Jackdaw connected to GitHub (or GitLab), and the Local Codebase source if you want the agent to see your files
MCP wired to your AI (Claude.ai, Cursor, or Claude Code)
The task your agent will work on

Step 1: Capture your constraints before you start

Tell the assistant the hard rules for this project, one at a time:

"For the migration project: never modify files in services/node/src/app/core/ directly — those changes go through PR review. Always run pnpm test before marking a task complete. Prefer small, atomic commits."

Each one is captured as a typed constraint in the project’s current state. From then on it is surfaced to the agent via get_context when relevant. You set the boundary once; the agent carries it across restarts.

Step 2: Write the project brief

In your wiki, type [[ and create a new page for this run as a Project — something like "Migration: Node → Python services." Fill it in with:

Goal — one sentence
Scope — what's in, what's out
Decisions made — choices already locked in
Progress — what's done, what's next

The agent can read this page with read_subject. The prose remains yours; the agent records durable changes as structured facts with remember.

Step 3: Start the agent with a grounded prompt

Your first prompt should make the agent read its context before doing anything:

You're working on a multi-day migration. Before you start:
1. Call get_context with "what is the migration project and its current state?"
2. Read the project brief.
3. Summarize what you know and what you're about to do.

Don't start work until you've confirmed your understanding.

Step 4: End each session with an update

Before you close a session, have the agent stage the durable changes it learned:

For each durable decision, constraint, or status change from this session,
call remember with the project name in the key and the precise value. Do not
edit the project brief. I'll review the proposed facts, then update my notes
if the narrative needs to change.

The proposed facts land in your inbox. Approve the accurate ones, and the next session can retrieve them alongside the human-authored brief.

Step 5: Spot-check for drift

At the start of each session:

Before we continue, tell me from your context (don't guess):
1. What constraints apply to this project?
2. What did we complete last session?
3. What's next?

If the answers are right, you're in sync. If something's off, correct it explicitly so your wiki reflects the truth.

Why this works

The agent may lose its conversational history at a reset, but it keeps access to Jackdaw. The Wiki preserves the project’s current state, the decisions and constraints behind it, and the evidence returned by earlier work. Day six starts from what the team settled on day one — plus everything useful learned since.