Most teams building agentic AI systems are moving faster than their testing practices. That gap is no longer theoretical. A 2026 safety benchmark found that none of 13 commercial AI agents cleared a 40% safe completion rate. Meanwhile, task-specific agents are expected to appear in 40% of enterprise software by the end of this year, up from less than 5% in 2025. The tools to evaluate them are still being written.
Evaluation Has Fundamentally Changed
Testing a language model and testing an agent are not the same thing. A model answers a question. An agent plans, selects tools, executes steps, and recovers from failures across a session that may span minutes or hours.
Adaline's 2026 benchmarking analysis frames this shift directly: the centre of gravity has moved from measuring model answers to measuring multi-step execution. Tool use, web navigation, file handling, terminal operations, and recovery from failed intermediate steps are now the units of assessment.
This has a practical consequence. Benchmarks now separate model capability from agent scaffold quality. Two deployments using the same underlying model can produce dramatically different results depending on how tools, memory, and prompts are wired together. Testing the model in isolation tells you very little.
The Safety Gap Is Measurable
NIST research from early 2025 found that novel attack strategies against AI agents achieved an 81% success rate in red-team exercises, compared to 11% against baseline defences. That is not a narrow gap. It reflects the expanded attack surface that agents carry by design: persistent memory, external tool access, and the ability to take consequential actions across systems.
Governance readiness is not keeping pace. A McKinsey AI Trust Maturity Survey in 2026 found that only 33% of enterprises meet governance standards for autonomous agents, with security cited as the primary barrier to scaling. Most organisations are deploying agents ahead of their own policy frameworks.
Standards Are Coming, But Teams Cannot Wait
The regulatory picture is not standing still. On 17 February 2026, NIST formally launched the AI Agent Standards Initiative, the first US government programme dedicated explicitly to interoperability and security standards for agentic systems. OWASP published its Top 10 for Agentic Applications in late 2025. The frameworks are taking shape.
The problem is timing. NIST's control overlays for single-agent and multi-agent systems remain in development. Teams deploying agents today are operating without finalised standards to reference.
That means defining your own evaluation criteria in the interim. At minimum: establish what safe task completion looks like for your use case, test how agents recover from failed tool calls, audit what data agents can read or write, and set explicit human review checkpoints for high-stakes actions.
EvalSpec is built to help QA teams and product managers apply this kind of structured evaluation to AI systems before they reach production.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation