ADLC Introduces 95% Success Benchmarks for AI Agents as Enterprises Add Control Planes
Updated
Updated · InfoWorld · Sep 24
ADLC Introduces 95% Success Benchmarks for AI Agents as Enterprises Add Control Planes
1 articles · Updated · InfoWorld · Sep 24
Summary
A proposed Agent Development Life Cycle reframes AI delivery around non-deterministic agents, requiring teams to define scope, boundaries and measurable targets before coding—such as handling 95% of valid requests without human help.
The framework extends software practice with continuous evaluation, observability, identity, tool-access controls and governance because agent quality depends on behavior during execution, not just final outputs.
OpenTelemetry-style semantics, runtime identities and policy enforcement are presented as key mechanisms for tracing tool calls, agent handoffs and model use while blocking unauthorized actions such as price changes or refunds.
An agent control plane would apply those controls across portfolios by registering agents, enforcing CI/CD quality gates, budgets and guardrails, and feeding production traces and online evaluations back into the next version.
The report argues organizations should build that foundation early, before multiple teams, frameworks and enterprise integrations create fragmented controls, weak visibility and higher security risk.
Will the massive bureaucracy of the new Agent Development Life Cycle destroy the very efficiency these autonomous AI systems were built to create?
If AI agents constantly rewrite their own execution paths, how can enterprise security teams truly guarantee these autonomous systems will never go rogue?
When an AI agent's unpredictable action looks like a mistake, how do evaluators know if it is a dangerous bug or a brilliant workaround?