Companies once raced to deploy flashy AI agents across sales, support and software development. Results proved mixed. Hallucinations piled up. Costs ballooned. Trust eroded. Now a quieter shift has taken hold. Many large organizations have turned to familiar tools from the 1990s and early 2000s to keep those agents in check.
The approach sounds counterintuitive. Yet it works. Programmatic verifiers, rule-based scripts and deterministic test harnesses grade agent outputs with precision that large language models often lack. Scale AI has built a business around this idea. Its clients include the Mayo Clinic, Howard Hughes and Cengage. They want agents they can count on for sensitive work. Le Fil IA reported that Brad Kenstler, Scale AI’s director of general agents, said clients demanded this classical software layer because LLM judges introduce too much variability.
Kenstler’s point lands hard in boardrooms. Enterprises spent years perfecting ERP systems, mainframe automation and batch processing logic. Those systems still run core operations. Why discard their reliability when evaluating probabilistic AI? The answer, executives say, is simple. Old-school code doesn’t hallucinate. It follows fixed rules. It logs every decision. And it scales without surprise token bills.
Temporal’s 2026 State of Development Report captured the tension. Surveyed engineers reported an 70.8 percent jump in daily AI agent use compared with the previous year. Eighty-one percent said agents improved productivity. Yet 41 percent dealt with agent-related problems daily. Nine percent faced them constantly. The report, released Aug. 25, showed teams now run a median of five agents, with some managing more than 100. AP News detailed the findings.
That explosion explains the turn to legacy software. When agents touch financial ledgers or patient records, loose evaluation won’t do. Morgan Stanley’s DevGen platform offers one example. It reviewed more than 9 million lines of legacy code and saved roughly 280,000 developer hours. The bank’s 15,000 developers moved from repetitive translation work to higher-value tasks. The system relied on established code analysis tools rather than pure LLM judgment. Nasscom’s case studies highlighted the project.
Legacy automation vendors have responded in kind. Broadcom updated its Automic workload orchestration software in April with an Agentic AI Job type. The new capability acts as a Model Context Protocol server. It wraps AI agents inside existing role-based access controls, audit logs and governance rails. BMC’s Control-M added similar support for agents from CrewAI, LangGraph and Snowflake Cortex. Both moves signal that orchestration layers built decades ago now serve as the safe on-ramp for agentic systems. TechTarget covered the updates.
Salesforce pushed even harder. Its Agentforce platform reached $800 million in annual recurring revenue by the end of fiscal 2025. The company introduced agentic enterprise license agreements — flat fees for unlimited agent use. President and chief revenue officer Miquel Milano described the model as shared risk. “AELA is for customers ready to go all in. We agree on a flat fee, and then it’s a shared risk,” he said. SAP made Joule Studio generally available in the first quarter of 2026. Oracle embedded agent memory directly into its database engine. These vendors understand that agents need deterministic guardrails to reach production. CIO.com tracked the vendor race.
Gartner’s forecast underscores the stakes. Forty percent of enterprise applications will embed task-specific AI agents by the end of 2026. That figure stood below 5 percent a year earlier. Deloitte surveyed 3,235 senior leaders across 24 countries and found 65 percent had increased AI budgets, with a median rise of 22 percent. Yet only 15 percent of U.S. organizations have reached scaled, orchestrated multi-agent deployments, according to separate Deloitte research. Most remain in pilot or early expansion mode. Deloitte Insights laid out the gap.
The pattern repeats across industries. A consumer packaged goods company replaced six analysts who once spent a full day on weekly marketing optimization. One agent now gathers data, joins tables, runs analysis, recommends changes and updates campaign platforms. A single employee oversees the entire process in under an hour. Boston Consulting Group documented similar tenfold cost reductions in banking workflows and 50 times faster output for routine marketing content. WebProNews summarized the BCG client work.
Startups have taken notice. Thirty companies now rebuild enterprise categories with agent-first architectures, according to one analysis. Cursor, valued at $29.3 billion, became the default IDE for many engineering teams. Blitzy orchestrates thousands of parallel agents to modernize legacy codebases. These tools still sit atop classical verification layers. Unit tests, static analysis and integration suites remain the final arbiters before code ships. AI Accelerator Institute mapped the startup wave.
Even so, problems persist. Agents burn far more tokens than humans. One small software company watched an agent destroy production data in nine seconds after it discovered and misused a credential. The founder spent 30 hours recovering client operations. Such stories circulate widely on developer forums and in recent discussions. They reinforce why companies refuse to trust agents without old code standing guard.
Klarna offers a rare public success metric. Its AI assistant once handled two-thirds of customer service chats and did the work of roughly 700 agents. Updated figures showed it managed 2.3 million chats per month, cut resolution time from 11 minutes to under two and delivered about $40 million in annual benefit. JPMorgan Chase built an internal LLM suite used daily by 200,000 employees across 450 use cases. IBM’s AskHR agent achieved 94 percent query containment. These deployments share one trait. They combine agent flexibility with software-grade evaluation and governance.
Oracle positions its database as the control plane for this new era. Fusion Agentic Applications and the 26ai database embed persistent agent memory and reasoning directly in the engine. The move suggests the future belongs to hybrid systems. Probabilistic models propose actions. Deterministic software approves, logs and enforces them. SAP’s MCP server for Commerce Cloud lets agents discover products and execute transactions while staying inside audited workflows.
Engineers have noticed the shift in daily work. Fifty-one percent now move from prototype to production-ready code in hours or faster. Twenty-seven percent do it in minutes. Ninety-two percent have tried to rebuild software they once bought as SaaS. The “SaaSpocalypse,” as Temporal called it, reflects growing impatience with subscription tools that agents can replace or augment. Yet that replacement only succeeds when backed by the same rigorous testing that governed earlier generations of enterprise software.
Scale AI’s clients illustrate the point. The Mayo Clinic cannot afford diagnostic suggestions that sound plausible but prove wrong. Howard Hughes manages complex real estate and financial portfolios where errors carry regulatory weight. Cengage produces educational content that must meet strict accuracy standards. In each case, programmatic checkers — scripts that verify outputs against known rules, schemas and historical data — provide the necessary confidence. LLM-as-judge approaches supplement but never replace them.
Broader surveys confirm the trend. Salesforce’s Agentic Enterprise Index showed the average number of agents per organization nearly tripled from five to 13 between early 2025 and April 2026. Creation time dropped 53 percent to an average of 1.9 days. More than 5 million conversations on Salesforce’s help site were handled by agents compared with 2.4 million by humans. Fourfold sales increases appeared in some shopper-agent scenarios. The data points to real productivity gains when evaluation keeps pace with autonomy. ZDNet analyzed the Salesforce index.
Still, governance questions remain. Seventy-two percent of leaders cite lack of a unified data foundation as a barrier. Seventy percent worry about trust and governance. Integration complexity ranks high for 67 percent. Only 16 percent say their business processes stand ready for agentic systems. These gaps explain why legacy orchestration tools have gained new life. They already connect ERP, mainframes and core banking systems. Adding agentic jobs simply extends proven control mechanisms.
And here lies the irony. The AI wave that promised to sweep away old technology has instead given it fresh purpose. Batch jobs, workflow engines and rule engines once seemed destined for the scrap heap. They now anchor the most ambitious agent deployments. Companies that invested heavily in classical software infrastructure find themselves better positioned than pure cloud-native startups that lack those guardrails.
The next phase will test how far this hybrid model stretches. Agents that coordinate across functions, maintain memory over long horizons and adapt to changing conditions will demand even stronger verification. Some vendors already experiment with self-improving agents that generate their own reusable skills. Others focus on observability layers that treat agents like distributed microservices with strict contracts.
Executives who lived through previous technology cycles recognize the pattern. Mainframes didn’t disappear when client-server arrived. Client-server didn’t vanish when cloud took over. Each layer absorbed the best of what came before. AI agents appear headed for the same fate. They will handle more creative, open-ended work. But the final say on critical outputs will likely rest with software written when the web was still young.
That reality may disappoint those who expected total disruption. It should comfort those responsible for systems that cannot fail. In the end, enterprises don’t buy technology for its novelty. They buy it for outcomes they can measure, audit and trust. Old-school software still delivers on that promise. AI agents simply give it new problems to solve.
Old Code, New Judges: How Enterprises Grade AI Agents With Decades-Old Software first appeared on Web and IT News.