AI Agent Platforms: Runtime Models, MCP, and Costs
A technical evaluation of developer AI agent platforms, covering container runtimes, MCP endpoints, compute units, and late-2026 model pricing increases.
Selecting an AI agent platform in late 2026 requires looking past UI builders to inspect runtime infrastructure and metering models. Production deployments diverge across container execution costs, Model Context Protocol (MCP) tool exposure, and scheduled model rate hikes rather than simple monthly seat licenses.123
Teams building enterprise workflows must balance infrastructure orchestration with raw token economics. Platforms like LangSmith meter execution via dedicated unit pricing, while cloud hyperscalers like Google Cloud and Microsoft Azure split fees across runtime environments and managed search grounding queries.234
Enterprise agent platform architecture in 2026
The developer platform ecosystem has bifurcated into orchestration engines and hyperscaler managed environments. Google Agent Platform integrates Google Antigravity to steer, customize, and orchestrate agents for workflows like product launches. Developers configure systems using the Agent Development Kit (ADK), accessing over 200 Google and third-party models and tools in Model Garden, including Gemini 4 Argon, the Claude model family, and open models like Gemma.4
Meanwhile, the legacy Amazon Bedrock Agents has formally entered maintenance mode under the name Amazon Bedrock Agents Classic and is closed to new customers. Existing Bedrock Agents Classic accounts retain the ability to run workloads without interruption, maintaining automated prompt engineering, memory, monitoring, encryption, permissions, and API invocation without custom code.1
Multi-agent frameworks have evolved from pure open-source codebases into managed infrastructure. LangGraph Platform provides dedicated hosting starting from $39 per user per month for engineering teams on cloud or self-hosted topologies. In parallel, CrewAI Enterprise delivers managed environments, live run monitoring, and RBAC team management, with integration toolkits that allow crews to invoke Amazon Bedrock Agents or existing CrewAI automations directly. Understanding these boundaries is central when building AI agents across varied production environments.56
Hosting models and tool orchestration: MCP, containers, and runtimes
Tooling protocols define how cleanly agents interface with internal databases and external APIs. Microsoft Foundry Agent Service supports three agent architectures: Prompt agents, Voice-based prompt agents, and Hosted agents. Prompt agents offer managed instructions and tool bindings; Voice-based prompt agents maintain persistent connections over a real-time WebSocket rather than discrete text requests; and Hosted agents run custom code packaged as container images or source code .zip files.7
Hosted agents in Foundry support frameworks including Agent Framework, LangGraph, OpenAI Agents SDK, Anthropic Agent SDK, GitHub Copilot SDK, or custom code. Developers curate tools including web search, file search, code interpreter, MCP servers, and custom functions, sharing them across agents via a single managed MCP endpoint with centralized authentication.7
LangSmith approaches tool interoperability from the serving side. LangSmith Deployment enables teams to expose an agent directly as an MCP server, supporting real-time streaming of intermediary steps and final outputs. When evaluating autonomous AI agents and their infrastructure, the runtime topology determines whether state management resides inside a managed container or across decoupled microservices.2
Platform pricing, seat plans, and compute metering
Platform operational fees separate developer seat access from runtime execution metering. LangSmith offers a Developer plan with 1 seat and up to 5k base traces per month at $0 before pay-as-you-go metering. The LangSmith Plus tier costs $39 per seat per month, including up to 10k base traces per month, deployment access, and 1 free Serverless Small deployment.2
Beyond base allowances, LangSmith meters deployment resources using LangChain Standard Units (LSU), priced at $1.00 per LSU. Base traces carry a 14-day retention limit, while extended traces store data for 180 days. Plus customers deploy on Cloud, while Enterprise tiers provide Cloud, Hybrid (SaaS control plane, self-hosted data plane), and fully self-hosted options.2
Microsoft Foundry Agent Service bifurcates billing based on agent type: Prompt agents charge per-call inference plus tool usage, while Hosted agents include per-call inference, tool usage, and container compute. CrewAI Enterprise delivers custom enterprise pricing covering cloud and self-hosted deployments. Hosted agents in Foundry can also be built directly with the GitHub Copilot SDK or Anthropic Agent SDK, where individual developer seats start at $10 a month for GitHub Copilot Pro or $25 a month for Claude Team Standard seats in our tracking data.578
Model token economics, scheduled step-ups, and timeout rules
Underlying LLM rates dictate agent iteration expenses. On Google Agent Platform, introductory pricing for Gemini 3.8 Flash, Gemini 3.7 Flash, Gemini 3.6 Flash, and CodeMender runs at $0.75 per 1M input tokens and $3.75 per 1M output tokens through December 31, 2026. On January 1, 2027, standard rates double to $1.50 per 1M input tokens and $7.50 per 1M output tokens.3
For flagship reasoning models, Gemini 4 Argon standard priority costs $2 per 1M input tokens globally ($0.10 cached) and $10 per 1M text output tokens globally through November 30, 2026. Starting December 1, 2026, global standard rates increase to $4 per 1M input tokens ($0.20 cached) and $20 per 1M output tokens.3
Execution billing mechanics contain strict operational constraints. Google Agent Platform bills only for requests returning an HTTP 200 response code, omitting 4xx and 5xx errors from input or output token charges. However, when a Gemini 4 completion exceeds 32K tokens in length, users are billed for retrievable token usage even if the request fails due to client or server timeouts, an update taking effect November 13, 2026. Detailed limits on Gemini 4 Argon architecture outline these long-output consumption boundaries.3
Grounding charges and context threshold penalties
Retrieval-augmented workflows accumulate secondary fees during multi-step tool calls. On Agent Platform, grounding with Google Search includes an allowance of 5,000 queries per month across Gemini 3 and Gemini 4 models, with excess queries billed at $14 per 1,000 queries. Grounding against private company datasets is billed at $2.50 per 1,000 prompts.3
Context window depth also triggers billing adjustments. If an input query context exceeds 200K tokens on Agent Platform, both input and output tokens are billed entirely at long-context rates. Managing state bloat in multi-turn agent loops is vital to prevent routine interactions from crossing this 200K threshold.3
Architectural selection hinges on tool exposure and deployment runtime needs. Teams relying on Python orchestrations with MCP endpoints benefit from LangGraph Platform or Foundry Hosted agents running containerized workflows. Teams standardizing on Gemini models must schedule budget revisions prior to the December 2026 and January 2027 token rate increases while implementing completion limits below 32K tokens to avoid timeout billing charges.237