The Warm-Up Problem: Why Your AI Agent Needs a Boot Sequence

Most AI agents start cold.

No context. No memory of what they were doing. No awareness of what they should never touch. Every session, they begin from scratch.

That's fine for simple tasks. It's a problem when you're running agents that operate autonomously across hours or days.

The fix is a boot sequence.

What a Boot Sequence Does

A boot sequence is a set of instructions that runs before the agent does anything else. It answers three questions:

1. Who am I and what's my job right now?
2. What state was I in when I last stopped?
3. What are my hard constraints?

Without this, agents make expensive mistakes in the first few steps of every session — misidentifying their scope, repeating completed work, or ignoring rules they should know cold.

The Minimal Boot Sequence

Add this to the top of your SOUL.md:

## Boot Sequence (run first, every session)
1. Read SOUL.md (identity + mission + constraints)
2. Read current-task.json (what am I working on?)
3. Read context-snapshot.json (where did I leave off?)
4. Check outbox.json (any unresolved escalations?)
5. Begin work

Five steps. Your agent starts every session informed instead of blind.

Why Current-Task and Context-Snapshot Are Separate

They solve different problems.

current-task.json answers: what am I supposed to do?

{
  "task": "review-and-post-daily-briefing",
  "scope": ["morning-content", "social-queue"],
  "do_not_touch": ["live-trading-positions", "email-drafts"],
  "started_at": "2026-03-09T10:00:00Z"
}

context-snapshot.json answers: where did I leave off?

{
  "last_completed_step": "drafted-7am-tweet",
  "next_step": "publish-newsletter",
  "notes": "Newsletter draft at content/drafts/march-9.md"
}

Merge these into one file and you lose the distinction between what you're doing and how far you've gotten. They have different lifecycles — the task might last a week while the snapshot updates every few minutes.

The Outbox Check

outbox.json is where your agent drops things it couldn't resolve. Escalations. Questions that need human input. Failed tasks it's not sure how to retry.

Checking it at boot means the agent resumes accountably. If there's an unresolved escalation from last session, it handles that first.

A five-second outbox check at boot has prevented broken publishes more than once.

What It Looks Like in Practice

We run 4 agents at Ask Patrick. Each has a boot sequence in its SOUL.md. When a session starts:

• The trading agent loads its current position context before touching anything market-related
• The content agent checks what was already posted so it doesn't double-publish
• The support agent checks the outbox for any escalated user questions from overnight

None of them start blind. That's not magic — it's a five-line boot sequence.

---

Full SOUL.md templates are in the Ask Patrick Library: askpatrick.co. New configs added every weekday.