When your AI agent hits an edge case, what does it do?
If you haven't explicitly defined escalation rules, the answer is: it guesses.
That's a problem.
What Escalation Rules Are
An escalation rule is a condition in your agent's config that says: "If X, stop and involve a human."
Without them, your agent will attempt to handle everything — including the situations where it absolutely shouldn't.
The Three Triggers Worth Configuring
1. Low-confidence writes
If the agent is about to write to a database, send an email, or post publicly, and its confidence in the output is below a threshold, it should pause and flag.
2. Financial outputs
Any output that touches money — invoices, expense reports, pricing quotes — should have a human in the loop until you've built significant trust.
3. First-time contact
When your agent reaches out to a person it has never interacted with before, a human review pass is worth the friction.
What the Config Looks Like
In practice, this is a small block in your SOUL.md or AGENTS.md:
## Escalation Rules
- Escalate to Patrick before sending any external email to a new contact
- Escalate before posting anything that mentions revenue, pricing, or customers
- Escalate if confidence in output is below 0.7 on any write operation
Ten lines. That's it.
Why Most Builders Skip It
Because it feels like overhead. You're building fast, the agent works in testing, and you want to ship.
But production is different. Edge cases multiply. The situations your agent handles gracefully in testing are not the situations that will break you — it's the ones you didn't think to test.
Escalation rules are the seatbelt. You don't notice them when things go right. You're glad they exist when things go sideways.
Running AI agents in production? The Ask Patrick Library has battle-tested configs for escalation, memory architecture, and cost control. askpatrick.co
Top comments (1)
The ten-line block in SOUL.md is the right vehicle — but the format inside it matters more than most people realize.
Prose escalation rules compete with everything else in SOUL.md. "Escalate before posting revenue details" sits next to role description, tone guidance, and behavioral notes. In context, they all have roughly equal weight. An agent reading SOUL.md as a document doesn't treat "escalate before X" as categorically different from "sound helpful."
The fix: a Constraints block with a dedicated escalation section, typed separately from the rest of the agent config. When constraints have their own block, they get treated as hard limits — the model's attention mechanism treats them differently than prose instructions buried in a role document.
The three triggers you list (low-confidence writes, financial outputs, first contact) are exactly the right taxonomy. What makes them stick in practice is making them structurally unambiguous — not mixed into the same section as "sound professional and concise."
flompt.dev / github.com/Nyrok/flompt