Catch the drift nobody else watches.
Most agent failures aren't sudden. They're slow capability creep + persona drift. An agent's behaviour gradually diverges from its declared system prompt over weeks. Runtime governance catches active bad behaviour — Drift-Observability catches the slow corruption.
What drift looks like
Real-world examples we've seen across customer fleets. Every one of them started clean and crept.
Tool-use creep
An agent meant for "lookup orders" gradually starts using "refund_customer" — first once a week, then daily. No single call looks wrong. The pattern does.
Output verbosity drift
A support agent that used to give 2-sentence answers starts giving 12-paragraph essays. Customer satisfaction drops 18% before anyone notices. Drift_score caught it.
Topic boundary shift
A finance agent starts answering relationship-advice questions. Within a quarter, 30% of its conversations are off-topic. Sentinel says: declared intent was "finance only".
EVOLVE-AGENT catches what — and what Drift-Observability catches differently
Runtime governance (EVOLVE-AGENT)
- Catches an agent doing something BAD right now
- Blocks a single destructive call · vetoes a quorum-required op
- Fires on every action · expensive at scale
- Misses gradual change because no single call is "bad"
- Designed for ATTACKS, not erosion
Drift-Observability (this product)
- Records a per-agent behaviour fingerprint over time
- Computes drift_score vs declared baseline · 0-100 scale
- Flags agents that cross threshold (default 25) for human review
- Cheap to run · scheduled scans, not per-action
- Designed for EROSION, not attacks
How drift gets caught
Lead records a behaviour fingerprint per agent per session. Compares it to the baseline. When the gap widens past your threshold, your team gets paged before a customer notices.
Observe
Fingerprint recorded every session — tool-call mix, response-length distribution, escalation rate, refusal rate. Lightweight, agent-side.
Compare
Continuous diff against baseline. Statistical drift score per axis. Trend tracked across days, weeks, months.
Review
Threshold crossed → your dashboard goes amber. Specific axis named. Sample sessions for review. Reset baseline if it's healthy drift; roll back deploy if it's regression.
Get Drift-Observability for your agent fleet.
Bundled into INTEGRITAS BUSINESS / SOVEREIGN · €1,500/mo as a standalone add-on.
Talk to us See INTEGRITAS tiers