AI Agents Are Moving Into Production, but Operating Costs and Controls Will Decide Their Value

What Happened

AT&T says it has nearly 500 agentic workflows in production, processes about 45 billion AI tokens a day, and plans to triple its workflow count next year [2]. NiCE is likewise positioning its contact-center agent platform around the difficulty of running agents, not merely building them [19]. Funding for WhiteMatter’s agent-oversight tools points to growing demand for controls over what agents access and do [17].

Model performance is becoming harder to separate from price. Google’s Gemini 4 Carbon took the lead on the MathArena benchmark, but reportedly costs about 15 times as much as GPT-6.1 Sol; it also exceeded the test’s time limit [8]. Anthropic made Claude Opus 5.5 the default on paid plans and advised users to revisit older prompting habits [9]. Separately, CEO Dario Amodei described compute purchasing as a major business uncertainty ahead of a planned IPO, with reported commitments approaching $901 billion [3].

In robotics, Runway released an open-weight, Apache 2.0-licensed video-trained controller that performs well on a robotics benchmark but cannot yet plan complex action sequences independently [14]. OpenAI is reassigning its standalone robotics researchers as it explores whether general-purpose models can take on more of that work [12]. At Nvidia, shareholders urged the board to consider Jensen Huang’s return amid reported leadership upheaval; a separate petition seeking his reinstatement drew substantial support [1][4].

Why It Matters to Businesses

AT&T’s scale makes the central question operational: can an agent complete a process reliably, within budget, and with accountable human oversight? A benchmark win does not answer whether a more expensive model improves enough real customer outcomes to justify its cost or latency [2][8]. Model and leadership changes also make dependence on one provider or open-model funding stream a procurement risk [6][9].

Kimbodo Engineering Perspective

We would treat agents as controlled workflow components, not autonomous replacements for whole departments. Use a capable model where judgment is necessary; use rules, smaller models, or conventional software for deterministic steps. The trade-off is less flexibility in exchange for predictable costs, testable behavior, and clearer failure ownership. Agent-monitoring products may help, but they should complement—not replace—permissions enforced by the underlying systems [17].

How We Would Implement It

• Choose one bounded workflow with a measurable baseline: completion rate, review time, error rate, latency, and cost per resolved case.
• Put agents behind an orchestration service that manages state, tool allowlists, retries, timeouts, and escalation. Give each tool the minimum permissions needed.
• Route tasks among models using evaluated quality, latency, and price; re-run task-specific tests when models or prompts change [8][9].
• Log model inputs, tool calls, approvals, and outcomes with appropriate redaction. Test against historical cases and adversarial scenarios before expanding autonomy.

Risks, Costs and Security

At high volume, small increases in tokens, retries, or model price can materially change unit economics [2][8]. Agents can also act on untrusted content, expose sensitive records, or make unauthorized changes. Require identity-scoped tool access, approval for consequential actions, audit trails, and a way to stop or roll back a workflow. For robotics and other physical systems, benchmark results are not a substitute for bounded operating conditions and independent safety controls [14].

Sources

1. [1] Letter: Altimeter's Brad Gerstner and Atreides's Gavin Baker urge Nvidia's board to consider "the potential return of Jensen Huang"; they own $2.13B in Nvidia (Katie Tarasov/CNBC)
2. [2] AT&T uses agentic AI to reinvent business processes from the ground up
3. [3] Anthropic CEO Amodei admits "there's a huge uncertainty in what's going to happen next year" days before a record IPO
4. [4] A petition to reinstate Jensen Huang as Nvidia CEO tops 100K signatures, including Elon Musk, Tim Sweeney, and Garry Tan; organizers say 5K+ Nvidia staff signed (Lauren Edmonds/Business Insider)
5. [6] Jensen Huang's resignation is a setback for US open models, as his funding gave open-source developers time to find their footing and will very likely dry up (Nathan Lambert/Interconnects AI)
6. [8] Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol
7. [9] Anthropic says outdated prompting habits are slowing down Claude Opus 5.5
8. [12] OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary
9. [14] Runway's open-source Praxis-1 turns video AI into a robot controller
10. [17] WhiteMatter, which builds "guardian" AI agents to oversee file access, software use, decision making, and actions of existing AI agents, raised a $20M Series A (Alice Tecotzky/Business Insider)
11. [19] NiCE’s next-generation Cognigy bets the hard part of AI agents is running them, not building them