700 engineering leaders are 74-77% confident in their AI agent controls, but the real safeguards trail by 30-55 points. See the confidence gap.

Engineering leaders are 74% to 77% confident in how they run AI agents in production. The controls that would back that up trail by 30 to 55 points. That's the AI agent confidence gap, and it comes from the Harness State of Agent DLC 2026 report, which surveyed 700 engineering leaders. The pattern held in all five areas we asked about: evals, security, inventory, cost, and rollback. Read the full report.
What the AI agent confidence gap looks like
We asked the same question five ways and got the same answer each time. Teams feel good about where they stand. The automated, always-on controls that would justify that feeling mostly aren't in place yet.
- Evals: 74% are confident their evals catch failures. Only 19% have a blocking gate on every release.
- Security: 75% are confident their agents are secure end to end. 87% had a security incident anyway.
- Inventory: 77% are confident they have no shadow AI. Only 44% run active discovery tooling to check.
- Cost: 74% are confident in their cost picture. 60% overran their AI budget.
- Rollback: 76% are confident they could roll back a bad agent within 15 minutes. Only 33% have a kill switch.
Confidence isn't evidence
Security shows the gap most clearly. Three in four teams believe their agents are secure end to end, and 87% have had a security incident. Not having a known incident is different from having visibility. A lot of that confidence probably reflects how long it's been since something visibly broke, not how well defended the system is.
Shadow AI works the same way. 77% say they have none, but only 44% run discovery tooling, so a lot of that 77% is a guess. Confidence with no way to detect problems behind it is just hope.
Three questions to ask your team
"Are we secure?" and "Do we have good evals?" almost always get a confident yes. Ask about the mechanism instead: what happens automatically, and what depends on a person noticing in time.
- If a bad agent release shipped right now, would something block it automatically, or would someone have to catch it?
- How would you know today if shadow AI was running somewhere you don't expect?
- Could you roll back a misbehaving agent in under 15 minutes? How sure are you, really?
These questions get past the "we're covered" reflex, and the gap between confidence and control usually shows up within a single conversation.
Where to start
Running agents on confidence instead of controls doesn't mean your team is behind. Most teams are in the same spot. Pick the area most likely to hurt you first: security if you're regulated, cost if budgets are tight, rollback if you're shipping fast. Then check whether the control there is automatic and blocking, or depends on someone catching the problem after the fact. You can close that gap well before it turns into an incident.
See how your controls compare. Download the State of Agent DLC 2026 report.



