Agent Infrastructure Goes Commodity as Prompt Tools Consolidate

The stack is now sold as a single subscription

On October 1, 2026, DigitalOcean introduced Agent Droplets, a product that puts the whole agent stack behind one subscription. Each Droplet bundles Firecracker microVM compute, memory, session storage, inference on hosted open models such as Kimi K3 and GLM 5.3, and governed access to more than 16,000 tools.22 There are three tiers: a $5 free trial, a $50-per-month Pro plan and a $200-per-month Team plan, each with unlimited agents and seats.24 The same day, Microsoft launched Agent 365 GCC. It puts agent discovery, Entra Agent ID, Purview data protection, Defender security and Intune device management into one FedRAMP High-authorized control plane for government clouds.24

On their own, these are product launches. Together they show a shift in how the market works. The runtime that hosts an agent, its sandbox, its memory and its tool connections used to be something engineering teams built and argued about. It is now becoming a line item you can buy. For anyone building or buying prompt management tools, evaluation suites or governance layers, the question is no longer whether you can run agents. It is which parts of the stack still let a vendor charge a premium.

Every vendor converged on the same design

The commodity case starts with how much the major platforms now look alike. In late September, AWS Strands, DigitalOcean Managed Agents and Aiven Runtime all shipped what Forkast describes as the same "harness" architecture: shell access, file manipulation, web tools, long-term memory and subtask delegation. OpenAI's DevDay then endorsed that pattern as the industry default.24

The hyperscalers' own documentation tells the same story. Amazon Bedrock AgentCore became generally available on October 13, 2025. It is a set of separate services (Runtime, Gateway, Memory, Identity, Browser, Code Interpreter and Observability), billed by usage and designed to work with any framework, including LangGraph, CrewAI and LlamaIndex.14 AWS rebuilt the runtime in September 2026, saying the new version releases memory as soon as sessions end and keeps cold starts near two seconds at any image size. The original version climbed toward 30 seconds on large images.12 AgentCore Evaluations, with 13 built-in evaluators, followed on October 15, 2026.15

Microsoft's Foundry Agent Service runs code built with Agent Framework, LangGraph, the OpenAI Agents SDK or the Anthropic Agent SDK as containers. Each agent gets a managed endpoint, autoscaling, its own Entra identity and saved session state.11 At Build, Microsoft said hosted agents would reach general availability within 30 days, with every session in its own sandbox and support for long-running autonomous agents.13 Google merged Vertex AI into what it now calls the Gemini platform in April 2026, renaming Agent Engine as Agent Runtime and adding a Managed Agents API in May.20

Three clouds now offer per-session isolation, accept any framework and charge by usage. One buyer's guide points out that the practical difference has little to do with capability and a lot to do with which cloud already holds your data and identity setup.15 That is about as clear a description of a commodity as you will find.

Price, not architecture, is now the competition

Once the technology looks the same everywhere, vendors compete on price. That started earlier than the October launches. In December 2025, Google cut Vertex AI Agent Engine runtime prices to get closer to AWS's preview rate of about $0.0895 per vCPU-hour. At the same time it added new charges for Sessions, Memory Bank and Code Execution starting in January 2026.30 AI CERTs called this textbook commoditization: the base compute gets cheap while add-on features carry the margin.30

DigitalOcean is betting on the other approach, which is a simpler bill. It says a typical agent team juggles six vendors and dozens of pricing units, so nobody can say what one agent run actually costs.22 AWS still bills memory, gateways and observability as separate metered items, while specialist sandbox providers such as E2B, Modal and Northflank compete on per-second billing.24 Forkast reads this split as a sign that the real differences between vendors have moved from engineering to accounting.22

The bundle has an important exception. Frontier models like GPT and Claude are not covered by the bundle discount and are billed at list price.22 My reading is that this matters more than the subscription itself. The infrastructure is getting cheap and predictable, but the intelligence it runs remains a variable cost set by the model labs.

Prompt management is being absorbed by bigger platforms

If the runtime is commoditizing from the bottom, prompt management tools are being absorbed from the top. Several independent observers describe the past year in this category as a shakeout. Humanloop shut down on September 8, 2025, after Anthropic hired its three co-founders and about a dozen engineers without buying the product or its IP.5 Mintlify acquired Helicone in March 2026, and the product now gets only security patches and new-model support.1 OpenAI bought Promptfoo the same month and folded it into its Frontier agent platform.2 ClickHouse acquired Langfuse in January 2026, alongside a $400 million Series D that valued ClickHouse at $15 billion.5 Vellum moved toward a consumer assistant in May 2026 and stopped publishing platform pricing.2 Palo Alto Networks completed its purchase of Portkey on May 29, 2026.5 Another tracker dates the announcement to April 30 and positions Portkey as the AI gateway for Prisma AIRS.8

The buyers say a lot. A database company bought Langfuse, and one analysis took that as evidence that AI observability is turning into core data infrastructure.21 A security company bought Portkey. A model lab bought Promptfoo. None of them wanted a standalone prompt store. They wanted versioning, tracing and red-teaming as features inside a larger product. Microsoft's Foundry already includes a preview "agent optimizer" that rewrites instructions, tool descriptions and model choices on its own.11 That is a prompt-management feature built into the runtime.

The Promptfoo deal shows best where the value is going. Futurum's Mitch Ashley argued that OpenAI was buying a way past the blocker that keeps enterprise deals stuck in evaluation, not adding a feature. He noted that Promptfoo was used at more than a quarter of Fortune 500 companies.6 Forbes reported that the startup had 11 people and $23 million in funding, and raised a conflict-of-interest problem: companies testing OpenAI models will now be using a tool owned by the vendor they are auditing.10 Reported user numbers also differ. Forbes cited more than 125,000 developers10, while Futurum counted more than 350,000 developers who had used its open-source tools.6 Either way, the reach was what OpenAI paid for.

The category is not shrinking for lack of demand. Braintrust raised an $80 million Series B at an $800 million valuation in February 2026.5 The independents that remain now advertise that they have no pending acquisition.8 Buyers are being told to check whether a vendor will survive before they sign.7

Where the moat is going, and where the analyses disagree

Most commentary agrees that frameworks, connectors and orchestration code are losing value. InformationWeek quoted practitioners advising companies to treat low-level orchestration as "a temporary advantage, not a permanent asset" and to invest in domain data, golden datasets, evaluation suites and governance policy instead.28 Turion makes the same point more sharply. It argues that MCP speeds up the race to zero for integration middleware, and that a vendor's moat grows with how much customer-specific state its service builds up over time.23

The disagreement is over which layer above the runtime wins:

• Turion ranks memory as the strongest moat, followed by governance, with execution sandboxes still emerging.23
• A widely shared Medium analysis names evaluation and governance as the most defensible layers, alongside vertical applications.21
• Forkast points to two things: the intelligence layer and compliance packaging. It notes that Microsoft's FedRAMP High, ITAR and CJIS certifications took years to earn, and it estimates that Agent 365 GCC saves federal agencies 12 to 18 months of custom compliance work.24
• Agentic AI Institute points to agent operating systems such as Sycamore, which raised a $65 million seed round, and to identity companies such as Keycard, which came out of stealth with $38 million.29

Taken together, the acquisitions give the clearest answer. Governance and evaluation are where buyers are actually spending money. OpenAI paid for red-teaming, Palo Alto Networks paid for a gateway with guardrails, and AWS shipped evaluations as a core service.6815 Memory may turn out to be the stickiest layer, but hyperscalers are already building it into their runtimes.1420 That leaves standalone memory vendors exposed to the same bundling that just reshaped prompt management.

Why it matters

The pressure behind all this is reliability, not capacity. Cisco research cited by Forkast found that only 5% of agentic AI projects have reached broad production.24 Analysts have warned that more than 40% of agentic projects could be cancelled by 2027.21 Commodity runtimes don't fix that. They take infrastructure off the list of excuses, so the remaining problems are testing, auditing and policy.

For builders, this suggests treating the runtime as replaceable and the evaluation data, policies and domain context as the assets to protect. For prompt-tooling startups, being bought means the market valued the feature, not the standalone company. With more users per vendor than ever, the risk now is getting bundled into someone else's platform before building a lasting business.