AI Agents Are Becoming Software Platforms: What Businesses Need to Secure Before Scaling

What Happened

• ChatGPT is moving toward an app platform. OpenAI announced plans to let users discover, launch and use software inside ChatGPT, potentially changing how applications reach customers. The announcement describes a direction, not a proven replacement for app stores. [3]
• Developer tools are becoming more extensible. Anthropic’s new Claude Code Mods let developers change the tool’s interface and behavior, including intercepting tool calls. [13]
• Agent governance is attracting investment. Reco raised $55 million to expand security and governance for agents operating across SaaS environments. Supabase raised $150 million and agreed to acquire Turso, which offers a database optimized for agents. [18][19]
• Production economics remain central. DoorDash described an internal GenAI platform serving more than 5,000 users, with LLM and agent gateways and a shift toward open-weights models. Nebius acquired Inferize, whose technology targets GPU utilization and inference costs. [7][16]
• Safety concerns remain unresolved. Former OpenAI safety employee David Robinson publicly criticized the company’s safety culture, citing incidents involving agents and internet-access restrictions. Separately, research on photo-to-3D agents found that tested systems were poor judges of their own reconstruction accuracy. [5][11]
• Compute supply chains are still a constraint. A reported inventory of Chinese fabs identified roughly 343 DUV immersion lithography tools by early 2026. Such equipment can be adapted to manufacture chips relevant to AI, though tool counts alone do not establish production capacity. [9]

Why It Matters to Businesses

AI adoption is shifting from standalone chat interfaces toward agents embedded in applications, developer workflows and operational systems. That increases the value of integrations, but it also gives model-driven software more opportunities to access data and take action. A convenient distribution channel or agent-ready database does not, by itself, provide authorization, auditability or dependable outcomes. [3][13][18][19]

Kimbodo Engineering Perspective

The practical decision is not whether to adopt agents wholesale. It is which actions can be delegated under which controls. DoorDash’s gateway approach illustrates the value of keeping model choice, cost and latency policies separate from each application. Robinson’s warnings and the 3D reconstruction findings point to a different limit: an agent’s confident account of its own behavior is not sufficient evidence that it acted correctly. [7][5][11]

How We Would Implement It

• Put an agent gateway between applications, models and tools. Route requests by task, sensitivity, latency and cost; log model versions, tool calls and outcomes. [7]
• Give each agent a scoped identity. Enforce permissions at the tool and data layers, with approval for consequential writes, external communications and privilege changes. [18][24]
• Run tool-using agents in isolated environments. Review extensions that can intercept calls, and test what happens when tools fail, return hostile content or exceed their intended scope. [13][25]
• Evaluate completed work against independent checks—such as database constraints, deterministic calculations or human review—rather than relying on agent self-assessment. [10][11]
• Track cost per successful task, not just cost per model request; compare hosted and open-weights options against accuracy, operations effort and utilization. [7][16]

Risks, Costs and Security

More integrations mean more credentials, data paths and failure modes. Gateway and approval layers add engineering work and some latency, while isolation and audit logs carry infrastructure costs. Those costs should be weighed against the impact of unauthorized actions or incorrect results. Start with read-only or reversible workflows, measure failures, and expand permissions only when independent tests show that the controls work. [18][23][25]

Sources

1. [3] OpenAI's DevDay 2026 announcements to turn ChatGPT into a place to discover, launch, and use software could potentially disrupt the traditional app store model (Sarah Perez/TechCrunch)
2. [5] Another OpenAI safety departure adds to a pattern of researchers leaving with public warnings
3. [7] Presentation: Building GenAI Platform at DoorDash
4. [9] Center for Technology & Statecraft: Chinese fabs had acquired ~343 DUVi tools by early 2026, with ~270 from ASML; DUVi can be adapted to make 7nm chips and HBM (Howard Liu/South China Morning Post)
5. [10] Open-source "BootLoops" harness supports AI models in performing precise scientific calculations
6. [11] AI agents build 3D scenes from photos but have no idea if they got it right
7. [13] Claude Code's new Mods system lets developers rewrite the AI coding tool from the inside
8. [16] Nasdaq-listed neocloud Nebius acquires Inferize, whose tech helps optimize GPU utilization and reduce AI request costs, sources say for $100M to $150M (Meir Orbach/CTech)
9. [18] Reco, whose tech helps enterprises secure and govern AI agents across SaaS environments, raised a $55M Series B extension, taking its total funding to $140M (Ram Iyer/TechCrunch)
10. [19] Supabase raised $150M led by Singapore's GIC and agrees to acquire Turso, which offers a database optimized for AI agents, for an undisclosed sum (Maria Deutscher/SiliconANGLE)
11. [23] CrowdStrike and CoreWeave bring AI security to machine speed
12. [24] NetApp hands storage operations to AI agents, but humans still draw the boundaries
13. [25] IBM and CoreWeave co-design controls for agent workloads