When Agents Run for Months and Spend Real Money, You Need More Than a Chatbot Budget

Perplexity shipped a digital employee last week that orchestrates 19 AI models and runs autonomously for months — priced at $200 a month. The same week, DBS Bank and Visa completed the first live agent-initiated credit card transactions in Asia Pacific, processing real food-and-beverage purchases on DBS/POSB credit and debit cards. And Fiserv became the first major payment processor to integrate both Visa's Trusted Agent Protocol and Mastercard's Agent Pay Acceptance Framework into its merchant acceptance systems.

Three announcements. Three different categories. One message: agents are no longer software features that respond and stop. They are persistent workers that run for months, spend real money at real merchants, and operate with genuine financial autonomy. McKinsey projects $3 to $5 trillion in agentic commerce by 2030. Gartner says 40% of enterprise applications will run task-specific AI agents by the end of this year — up from less than 5% in 2025. Microsoft reports that 80% of Fortune 500 companies already have active agents in production. But most of those deployments were designed for session-based interactions. An agent does a task, a human reviews, done. The infrastructure was built for chatbots, not permanent digital employees.

This is where the gap kills projects. Platforms like AgentPMT exist because this moment was inevitable — when agents need agent wallets with programmable budget controls, Dynamic MCP for on-demand tool governance with zero context bloat, and drag-and-drop workflow builders that define controlled multi-step processes. The businesses treating agents like enhanced chatbots — no budgets, no tool governance, no financial audit trails — are the 40% Gartner predicts will scrap their agentic projects by 2027. Not because the models failed. Because the operational infrastructure never existed.

The Permanent Agent Is Here

Perplexity Computer launched February 25-27 and represents something fundamentally new: the first consumer-accessible persistent autonomous agent platform. It orchestrates 19 specialized AI models — Opus 4.6 for core reasoning, Gemini for deep research and sub-agent creation, Grok for speed on lightweight tasks, ChatGPT 5.2 for long-context recall, Nano Banana for image generation, Veo 3.1 for video. The system creates sub-agents autonomously, finds API keys, researches information, programs applications when needed, and contacts the user only when truly necessary.

At $200 per month on the Max tier, persistent autonomous agents are now accessible to individual users and small teams — not just enterprise. Perplexity followed up on February 28 by announcing an enterprise expansion with usage-based pricing and per-task spending caps. The company emphasized that model specialization, rather than commoditization, is shaping the AI landscape and signaled its priority is enterprise subscriptions and revenue growth over mass-market user metrics.

This is fundamentally different from chatbot-era deployment. You do not invoke a persistent agent and wait for a response. You hire it, manage it, and govern it. When those agents run for months and spawn sub-agents dynamically across 19 models, the question of operational infrastructure stops being theoretical. AgentPMT's agent wallets on the Base blockchain provide the financial governance persistent agents demand: budget controls at daily, weekly, monthly, or per-transaction granularity, a full on-chain audit trail mirrored in the dashboard, and stablecoin denomination for predictable costs. The credit system — 100 credits equal $1 USD, only charged on successful tool calls — means you pay for results, not attempts. When an agent runs for months and autonomously spawns sub-agents, granular financial control is the difference between managed operation and a runaway expense report.

The Payment Rails Are Live

DBS Bank and Visa completed the first live agent-initiated credit card transactions in Asia Pacific on February 16, processing real food-and-beverage purchases through Visa Intelligent Commerce. The pilot uses AI-ready credentials — secure tokenized card details for trusted agents — with advanced authentication and intent-driven transaction controls. DBS, Southeast Asia's largest bank operating across 19 markets, is the first issuer in the region to advance real-world agentic commerce. In Singapore, 77% of residents already use generative AI chatbots daily and 80% rely on AI assistance when shopping online. The consumer behavior is there. The banking infrastructure just caught up.

On the merchant side, Fiserv became the first major processor to integrate both Visa's Trusted Agent Protocol and Mastercard's Agent Pay Acceptance Framework. Sanjay Saraf, Fiserv's global chief product officer for merchant solutions, stated that the companies are "working together to establish the foundation for secure, intelligent and interoperable agentic commerce experiences." The integration uses network tokenization that replaces card numbers with network-issued tokens, authenticates authorized AI agents while filtering malicious automation, and routes transactions through standard authorization, settlement, and reconciliation flows. Merchants on Fiserv's Clover platform can accept agent-initiated transactions without building custom logic.

McKinsey's analysis of the agentic commerce opportunity highlights a critical shift for merchants: brand storytelling and front-end experience matter less to agents than operational trust. Agents optimize for delivered value — price, availability, service reliability, reversibility. Merchants that expose clean inventory data and transparent policies become default suppliers. The ones that do not become invisible to AI-mediated commerce.

From the agent operator's side, controlling what agents spend across both traditional card rails and crypto-native rails is its own infrastructure challenge. AgentPMT's agent credit card integration handles this by injecting stored payment credentials server-side at the moment of purchase — the agent initiates a transaction, AgentPMT securely provides the credentials, and the agent receives only a confirmation. The agent never sees the card number, CVV, or expiration date. Combined with agent wallets for crypto-native payments via x402 and x402Direct on Base, AgentPMT provides unified financial infrastructure across both payment methods with budget controls on top of both.

AgentOps: The Discipline That Just Became Mandatory

IBM now formally defines AgentOps as the set of practices, tools, and frameworks used to design, deploy, monitor, optimize, and govern autonomous AI agents in production — positioning it alongside DevOps and MLOps as a required operational discipline. IBM estimates the AgentOps market at $5 billion in 2024, projecting growth to approximately $50 billion by 2030. AWS Marketplace already offers enterprise AgentOps solutions. The category is maturing from concept to purchasable product.

The data on enterprise readiness is both encouraging and alarming. KPMG's Q4 AI Pulse survey of C-suite leaders at billion-dollar organizations found that 75% prioritize security, compliance, and auditability as the most critical requirements for agent deployment. Sixty percent restrict agent access to sensitive data without human oversight. Nearly half employ human-in-the-loop controls across high-risk workflows. But 80% now cite cybersecurity as the greatest barrier to their AI strategy — up from 68% at the start of 2025. And 65% cite agentic system complexity as the top operational barrier, a number that has held for two consecutive quarters. The agentic AI security crisis is accelerating faster than most organizations can respond.

Microsoft's Cyber Pulse report confirmed that 80% of Fortune 500 companies use active AI agents, framing observability, governance, and security as the three critical pillars for enterprise adoption. Help Net Security reported that 30% of organizations now have dedicated AI security budgets, up from 20% the prior year, with half of executives planning to allocate $10 to $50 million to secure agentic architectures. Spending is real. The question is whether it maps to actual operational infrastructure or just more monitoring dashboards bolted onto chatbot-era systems.

Gartner's prediction carries the sharpest edge: 40% of agentic AI projects will be scrapped by 2027. Not because the models fail — because organizations struggle to operationalize them. Trace, which raised $3 million from Y Combinator on February 26, is attacking one piece of the puzzle: mapping complex corporate environments so agents understand what they are working with. But environment mapping is one component. The full AgentOps stack requires budget management per agent and per task, tool access governance with the principle of least privilege, multi-model orchestration, financial audit trails for every transaction, human-in-the-loop checkpoints for high-risk operations, and workflow definitions that constrain agent behavior to approved paths.

AgentPMT provides this full stack as a single platform. Dynamic MCP ensures agents only access approved tools, fetched on-demand with zero upfront context consumption — no tool bloat, no wasted tokens. The drag-and-drop workflow builder defines controlled multi-step processes with clear success criteria at every step. Every tool call has a visible price. Every workflow has a total cost. Human-in-the-loop via the AgentPMT mobile app lets agents request human approval during any workflow step. And every agent interaction is logged with full request/response capture, workflow step tracking, prompt correction capabilities, persistent sessions, and compliance-ready audit trails.

What This Means for You

The shift from agent-as-tool to agent-as-worker creates an infrastructure category that every business deploying AI agents must address now, not next quarter. The businesses that treat agent deployment like adding a software feature will be Gartner's 40%. The businesses that build operational infrastructure — budgets, tool governance, financial controls, audit trails, multi-model management — will capture the $3 to $5 trillion McKinsey projects. Operating an agent fleet demands the same rigor as managing a distributed workforce.

Perplexity Computer is running agents for months right now. DBS Bank agents are spending real money on credit cards right now. Fiserv merchants can accept agent-initiated payments right now. AgentPMT was built for exactly this inflection point: model-agnostic operation across every LLM — including the 19 models Perplexity Computer just validated as the right architecture — agent wallets with programmable budget controls at every granularity, the largest marketplace of AI tools and skills accessible via Dynamic MCP, and a workflow builder that turns complex multi-step processes into controlled, auditable operations.

What to Watch

The NIST March 9 deadline for AI agent security comments through the CAISI initiative may shape formal agent governance requirements and compliance standards. The Commerce Department's March 11 evaluation of state AI laws could affect how agent deployment is regulated across jurisdictions, creating compliance complexity for multi-state operations. Whether Fiserv's dual-network integration becomes the standard for merchant agent acceptance depends on whether other major processors follow quickly. Perplexity Computer's enterprise rollout will determine whether months-long autonomous operation goes mainstream. And the AgentOps vendor landscape will start sorting winners from also-rans as the category matures from definition to deployment.

The chatbot era taught businesses to think about AI in terms of conversations. The agent era demands they think in terms of operations — budgets, governance, compliance, financial controls, tool access, model management. Every agent that runs longer than a single session is a worker that needs managing. The payment rails are live. The multi-model orchestration works. The question is not whether your agents are capable enough — it is whether your infrastructure can govern what they are about to do.

Explore AgentPMT's agent operations infrastructure →

Key Takeaways

• Agents that run for months and spend real money require operational infrastructure — budgets, tool governance, financial controls, audit trails — that most businesses have not built yet, and the window to build it is closing fast
• Gartner predicts 40% of agentic AI projects will be scrapped by 2027 because organizations cannot operationalize them, not because the models fail
• The payment rails for agent-initiated commerce are consolidating rapidly, with Fiserv integrating both Visa and Mastercard agent frameworks and DBS completing live transactions in Asia Pacific

Sources

• Perplexity launches Computer AI agent that coordinates 19 models, priced at $200 a month — VentureBeat
• DBS is first bank in Asia Pacific to pilot Visa Intelligent Commerce for everyday payments — DBS Newsroom
• Fiserv Integrates Mastercard Agent Pay Into Merchant Platform — PYMNTS
• Fiserv, Visa unite on agentic tools — Payments Dive
• The agentic commerce opportunity — McKinsey
• Agentic Commerce Success Centers On Overcoming Key Frontiers In Consumer Trust — Forrester
• What is AgentOps? — IBM
• Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 — Gartner
• AI at Scale — KPMG
• 80% of Fortune 500 use active AI Agents — Microsoft Security Blog
• AI Agent Payment Solutions in 2026 Compared — Privacy.com
• AI risk moves into the security budget spotlight — Help Net Security
• Trace raises $3M to solve the AI agent adoption problem — TechCrunch
• Perplexity Unveils Enterprise-Focused AI Agent System — The AI Insider