AI Agents Are Gaining Users Faster Than Reliability and Infrastructure Can Keep Up

What Happened

DeepSeek says its open-source agent app Harness reached 1.54 million daily active users less than three months after launch. It also reports processing roughly 92 trillion tokens a day—about 60 million per user. Those are company-reported usage figures, not a measure of completed work or business value. [1] Apple, meanwhile, is reportedly pushing to move beyond Siri improvements toward a more capable AI agent. [3]

Evidence of the execution gap remains stark. In Alibaba’s Co-Design Bench, which uses real coding requests that previously challenged agents, the top model fully solved 26.4% of tasks, despite scores above 99% on older benchmarks. [9] Outside software development, Andrej Karpathy used Claude to build a curated video platform in a weekend, illustrating how quickly an individual can assemble a useful application when its output can be inspected. [5]

Capacity is becoming a strategic constraint. Anthropic reportedly committed $3.4 billion for 1.7 GW of gas-fired generators, while AI companies have sold or priced more than $300 billion of stock this year, excluding Anthropic’s IPO, with roughly $100 billion more planned. [4][2] AI oversight also became a significant electoral issue: an Anthropic-backed safety group spent more than $63 million in the midterms, and 10 of its 17 supported congressional candidates won. [14]

Why It Matters to Businesses

Agent adoption signals demand, but usage and benchmark scores are poor substitutes for measuring whether an agent finishes a business task correctly. Buyers should budget for review, retries and escalation—not assume that a capable model makes an autonomous workflow reliable. At the same time, power procurement and capital requirements could affect model pricing, capacity and supplier concentration. [1][9][4][2]

Kimbodo Engineering Perspective

The practical distinction is between an agent that produces a plausible answer and a system that completes an authorized task with a verifiable result. A coding agent’s output can be tested; a customer-facing agent may need policy checks, transaction limits and human approval before it acts. Greater autonomy is worthwhile only where the time saved exceeds the cost of failures, supervision and inference. The benchmark gap argues for testing on an organization’s own difficult cases rather than selecting a model from a leaderboard. [9]

How We Would Implement It

• Start with one bounded workflow, a defined success measure and a baseline for human time and error rates.
• Build an evaluation set from real requests, including ambiguous instructions, failed cases and security-sensitive actions; score completed outcomes, not just responses.
• Give the agent narrowly scoped tools. Require explicit approval for irreversible actions, and record tool calls, inputs, outputs and reviewer decisions.
• Route routine tasks to lower-cost models where they pass the same tests; use stronger models or human review when confidence or risk thresholds require it.
• Track cost per successfully completed task, including retries and review, alongside latency and failure severity.

Risks, Costs and Security

Long-running agents can multiply token costs and expose more data through tool use; the scale of Harness’s reported token consumption makes per-task accounting essential. [1] Prompt injection, excessive permissions and unreliable outputs require least-privilege access, isolated execution where appropriate, audit logs and clear recovery paths. Infrastructure planning should also account for supplier availability and energy exposure, rather than treating inference capacity as unlimited. [4] Governance requirements may change as AI oversight gains political weight. [14]

Sources

1. [1] DeepSeek says its open-source agent app Harness hit 1.54M DAUs on November 4, less than three months after launch, and uses ~92T tokens a day, or ~60M per user (Saritha Rai/Bloomberg)
2. [2] Excluding Anthropic's IPO, AI companies sold or priced $300B+ of stock in 2026, with ~$100B more on tap, raising questions about how much the market can absorb (Alex Eichenstein/The Information)
3. [3] Sources: John Ternus wants more urgency in Apple's AI efforts, pushing Craig Federighi to move beyond Siri AI tweaks to agentic and other advanced capabilities (Mark Gurman/Bloomberg)
4. [4] Sources: Anthropic bought commitments for 1.7GW of gas-fired generators from equipment supplier OnePWR for $3.4B, the largest known such purchase by an AI lab (Martha Muir/Financial Times)
5. [5] Andrej Karpathy built a curated video platform in a weekend using Claude and 20 years of Hacker News
6. [9] AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
7. [14] Anthropic-backed AI safety super PAC Public First spent $63M+ in the 2026 midterms, and 10 of the 17 congressional candidates it backed won their elections (Emily Birnbaum/Bloomberg)