Bedrock AgentCore
Pricing, Core Services, and a 2026 Setup Guide for AWS's Agent Runtime
Amazon Bedrock AgentCore is AWS's managed platform for deploying, running, and securing AI agents in production, bundling seven composable services (Runtime, Gateway, Memory, Identity, Browser, Code Interpreter, and Observability) behind consumption-based pricing. Since reaching general availability on October 13, 2025, it has become one of the fastest-searched terms in AWS's AI portfolio, and this guide covers exactly what it costs, how each service works, and how it stacks up against Azure AI Foundry and Vertex AI Agent Builder.
Key Takeaways
- 2,900 searches/month for "bedrock agentcore" reflect fast-growing developer demand and low keyword difficulty (13)
- Runtime pricing is $0.0895 per vCPU-hour and $0.00945 per GB-hour, billed only while your agent is actively working
- 8-hour execution windows and full session isolation make Runtime suited to long, asynchronous agent tasks
- Gateway turns existing APIs, Lambda functions, and MCP servers into agent-ready tools for $0.005 per 1,000 calls
BEDROCK AGENTCORE AT A GLANCE (2026)
What Is Amazon Bedrock AgentCore?
Bedrock AgentCore is AWS's answer to the "last mile" problem of AI agents: it is easy to prototype an agent locally, but hard to run it securely, at scale, for real users. AgentCore packages the infrastructure that production agents actually need — isolated compute, tool access, memory, identity, and monitoring — as separate services you can adopt individually or together.
Unlike Bedrock's original Agents feature, AgentCore is framework-agnostic. According to the official AWS launch post, it supports agents built with LangGraph, CrewAI, Strands Agents, LlamaIndex, or raw code, and can invoke any model — Bedrock-hosted or external — rather than locking you into a single orchestration layer.
Why AWS Built AgentCore
Most teams that shipped an agent prototype in 2024 and 2025 hit the same three walls: agent sessions outlive Lambda's 15-minute timeout, connecting agents to internal tools means writing brittle custom adapters, and there is no standard way to audit what an autonomous agent actually did. AgentCore addresses all three with dedicated services rather than workarounds bolted onto existing compute products.
Framework & Model Agnostic Design
AgentCore hosts agents on top of Amazon's marketplace of 200+ models from Anthropic, Meta, Mistral, AI21, and others, and can also call out to external APIs and non-Bedrock models. This is the trait that separates it most clearly from Azure and Google's equivalents, both of which nudge developers toward their own first-party model (Azure OpenAI models, Gemini) as the default reasoning engine.
What AgentCore Solves for Production Agents
Core AgentCore Components
AgentCore is not a single product but seven services you can mix and match. Here is what each one actually does.
1AgentCore Runtime
A serverless execution environment purpose-built for agent workloads. Runtime supports execution sessions of up to 8 hours with complete session isolation between users — each session runs in its own dedicated micro-VM with isolated CPU, memory, and filesystem. At general availability, Runtime added support for the Agent-to-Agent (A2A) protocol, letting agents delegate subtasks to other agents.
Session Isolation in Practice
Because each user gets a fresh, isolated sandbox, one user's agent session cannot read another's memory, credentials, or file state — a meaningful advantage for multi-tenant SaaS products handling sensitive customer data.
2AgentCore Gateway (MCP Gateway)
Gateway converts existing REST APIs, OpenAPI specs, and AWS Lambda functions into Model Context Protocol (MCP)-compatible tools, and can also connect directly to existing external MCP servers via the SynchronizeGatewayTargets API. It indexes every available tool for semantic search, so an agent can discover the right tool from a natural-language description rather than a hardcoded tool list.
Turning Internal APIs Into Agent Tools
For teams that already run internal REST services or Amazon API Gateway endpoints, this is the fastest path to agent-enabling existing infrastructure without rewriting a single backend service.
3AgentCore Memory
Memory gives agents recall across sessions. Short-term memory stores raw event streams within a session; long-term memory automatically extracts and consolidates facts, preferences, and summaries for retrieval in future sessions. At GA, AWS added a self-managed strategy giving teams full control over their own extraction and consolidation pipelines instead of relying solely on the built-in one.
4AgentCore Identity
Identity manages how agents authenticate to other services on a user's behalf. It offers secure vault storage for OAuth refresh tokens, identity-aware authorization, and native integration with OAuth-enabled third-party services — so an agent can, for example, act inside a user's Google Calendar or Slack without the application ever holding a raw password.
5Browser Tool & Code Interpreter
Browser gives agents a managed, sandboxed headless browser for web tasks (form filling, research, checkout flows). Code Interpreter runs agent-generated Python or JavaScript in an isolated sandbox for data analysis and calculations, billed on the same vCPU/GB-hour model as Runtime.
6Observability
Observability streams traces, reasoning steps, tool calls, and latency into Amazon CloudWatch, billed under standard CloudWatch pricing rather than a separate AgentCore rate. This is what makes agent behavior auditable after the fact instead of a black box.
Also worth knowing: as of December 2025, AWS added Policy (guardrail rules for agent behavior) and Evaluations (automated agent quality scoring) as preview services, alongside a Payments and Search capability — bringing the full AgentCore family closer to twelve billable components. Expect these to matter more once they exit preview.
Bedrock AgentCore Pricing Breakdown (2026)
AgentCore uses pure consumption-based pricing: no seats, no platform fee, no minimum commitment. You only pay for active compute time, API calls made, and memory stored or retrieved. Here is every published rate as of the current AWS pricing page.
| Service | Meter | Price |
|---|---|---|
| Runtime / Browser / Code Interpreter | Active vCPU-hour | $0.0895 |
| Runtime / Browser / Code Interpreter | Peak GB-hour | $0.00945 |
| Gateway | Per 1,000 API invocations | $0.005 |
| Gateway | Per 1,000 search invocations | $0.025 |
| Gateway | Per 100 tools indexed / month | $0.02 |
| Memory (short-term) | Per 1,000 new events | $0.25 |
| Memory (long-term, built-in) | Per 1,000 records stored / month | $0.75 |
| Memory (long-term, self-managed) | Per 1,000 records stored / month | $0.25 |
| Memory (retrieval) | Per 1,000 record retrievals | $0.50 |
| Identity | Per 1,000 token/API-key requests | $0.010 |
| Observability | Standard CloudWatch rates | Varies by region |
Source: AWS Bedrock AgentCore pricing page. Identity is free when used exclusively through Runtime or Gateway. Rates exclude the underlying Bedrock model token costs, which bill separately per your chosen model.
Where the Bill Actually Comes From
Cost analysis published by AWS notes that because agents typically spend 30 to 70 percent of a session idle — waiting on an LLM response or an external tool call — idle CPU time is not billed, which meaningfully lowers Runtime costs versus naive always-on compute. In practice, Runtime duration and long-term Memory storage are the two dominant line items for most production agents, not Gateway or Identity.
Example Monthly Cost Calculation
Consider a mid-size support-automation agent handling roughly 4,000 conversations a month, each averaging 6 minutes of active compute:
Sample workload: 4,000 conversations/month
- Runtime: ~400 vCPU-hours + ~800 GB-hours → $35.80 + $7.56 = $43.36
- Gateway: ~600,000 tool calls → $3.00
- Memory: ~250,000 short-term events + 40,000 long-term records → $62.50 + $30.00 = $92.50
- Identity: ~20,000 token requests → $0.20
- Estimated infrastructure total: roughly $139/month, excluding model token spend and CloudWatch
Note: Model inference tokens (the actual Claude, Nova, or Llama calls) are typically the largest overall cost and are billed separately under standard Bedrock model pricing.
Interactive Bedrock AgentCore Cost Estimator
Model your own workload against current AWS list prices. Adjust the sliders to match your expected usage and see the estimated monthly Runtime, Gateway, and Memory bill update instantly.
Bedrock AgentCore Cost Estimator
Drag the sliders to model a monthly Runtime, Gateway, and Memory bill using current AWS list prices. Estimates only — actual invoices also include Observability (CloudWatch) and model token spend.
Bedrock AgentCore vs Azure AI Foundry vs Vertex AI Agent Builder
All three hyperscalers now ship a managed agent runtime. The meaningful differences are in model neutrality, ecosystem lock-in, and where your data already lives.
| Dimension | Bedrock AgentCore | Azure AI Foundry Agent Service | Vertex AI Agent Builder |
|---|---|---|---|
| Model neutrality | Highest — 200+ models, any framework | Azure OpenAI-first, 1,700+ catalog entries | Gemini-first by default |
| Session length | Up to 8 hours, isolated micro-VMs | Threads persisted in Cosmos DB | Session state in Memory Bank |
| Identity integration | OAuth vault + IAM | Deep Entra ID binding | Google Cloud IAM |
| Best fit | AWS-native teams, custom frameworks | Microsoft 365 / Teams-heavy orgs | BigQuery-anchored, Google Search grounding |
| Pricing model | Pure consumption, no minimums | Consumption + Azure OpenAI token rates | Consumption + Gemini token rates |
| Notable scale limit | ~100 concurrent sessions/region by default | Quota-based, region dependent | Quota-based, region dependent |
When to Choose Each Platform
Choose Bedrock AgentCore if:
You are already on AWS, want to avoid committing to one model vendor, or need to bring an existing LangGraph/CrewAI agent into a managed runtime without rewriting it.
Choose Azure AI Foundry Agent Service if:
Your organization runs on Microsoft 365, Teams, and Entra ID, and you want agents that plug directly into that identity and collaboration layer.
Choose Vertex AI Agent Builder if:
Your data already lives in BigQuery and GCS, or your agent needs first-party Google Search grounding and Gemini's native multimodal reasoning.
For a deeper look at how these three stack up against smaller cloud-native agent platforms, see our cloud AI agent platforms comparison.
How to Get Started with Bedrock AgentCore
A minimal path from zero to a deployed agent:
- Pick or bring a framework: Use Strands Agents (AWS's own SDK), or bring an existing LangGraph, CrewAI, or LlamaIndex agent.
- Deploy to Runtime: Package your agent and deploy it to a Runtime endpoint; each invocation gets its own isolated session.
- Wire up tools via Gateway: Point Gateway at your internal APIs, Lambda functions, or existing MCP servers so the agent can discover and call them.
- Add Memory: Enable short-term memory for in-session recall and long-term memory for cross-session personalization.
- Configure Identity: Set up OAuth connections for any third-party service the agent needs to act on behalf of a user.
- Turn on Observability: Route traces to CloudWatch before going to production, not after an incident.
AWS also publishes an AgentCore MCP Server with built-in scaffolding for Runtime, Gateway, Identity, and Memory, which shortens initial setup considerably if you are starting from scratch. For general MCP concepts and other gateway options, our complete Model Context Protocol guide, MCP gateway comparison, and AWS MCP servers guide cover the protocol in more depth.
Limitations & Pitfalls
Where Planetary Labour Fits Into This Picture
Bedrock AgentCore is infrastructure — it gives engineering teams the building blocks to run agents in production, but somebody still has to design, prompt-engineer, monitor, and iterate on what those agents actually do. Planetary Labour sits one layer up: it is an autonomous GTM engine that runs the marketing work itself — publishing SEO content, posting to X and Reddit, and building domain authority — without your team having to operate the underlying agent infrastructure at all.
If you are evaluating AgentCore to build an internal agent platform, that is a good sign your organization is ready for agentic systems generally. For the specific job of growth and distribution, most founders would rather not stand up Runtime, Gateway, and Memory themselves — that is exactly the gap an always-on GTM engine like Planetary Labour is built to close, while your engineers focus AgentCore-style infrastructure on the product itself. Read more in our guide to enterprise AI agent adoption.
Frequently Asked Questions
What is Amazon Bedrock AgentCore?▼
Amazon Bedrock AgentCore is a set of composable AWS services — Runtime, Gateway, Memory, Identity, Browser, Code Interpreter, and Observability — for deploying and operating AI agents securely at scale. It reached general availability on October 13, 2025, and works with any model or agent framework, including LangGraph, CrewAI, Strands Agents, and LlamaIndex.
How much does Bedrock AgentCore cost?▼
AgentCore is consumption-based with no minimum fees. Runtime costs $0.0895 per vCPU-hour plus $0.00945 per GB-hour. Gateway is $0.005 per 1,000 API invocations. Short-term Memory is $0.25 per 1,000 events, and long-term Memory storage is $0.75 per 1,000 records per month. Observability bills separately through Amazon CloudWatch.
What is the AgentCore Gateway used for?▼
AgentCore Gateway converts existing APIs, AWS Lambda functions, and OpenAPI specs into agent-callable tools exposed over the Model Context Protocol (MCP). It also connects directly to existing MCP servers, indexes tools for semantic search, and centralizes authentication so agents can discover and call tools without custom integration code.
How long can an AgentCore Runtime session run?▼
AgentCore Runtime supports execution sessions of up to 8 hours with complete session isolation between users, plus Agent-to-Agent (A2A) protocol support at general availability. This makes it suited to long-running, asynchronous agent workloads that traditional Lambda timeouts cannot handle.
Is Bedrock AgentCore better than Azure AI Foundry or Vertex AI Agent Builder?▼
It depends on your cloud footprint. AgentCore is the most framework- and model-agnostic of the three and fits AWS-native teams best. Azure AI Foundry Agent Service wins for deep Microsoft 365 and Entra ID integration. Vertex AI Agent Builder is strongest when you need Gemini grounding and tight BigQuery data gravity.
Next Steps with Bedrock AgentCore
You now have the full pricing model, the seven core services, and a decision framework against Azure and Google's equivalents. A sensible next step:
- 1. Run the cost estimator above with your real expected volumes before committing to an architecture.
- 2. Prototype with Runtime + Gateway only — add Memory and Identity once the core loop works.
- 3. Read the official docs: the AgentCore developer guide covers setup end to end.
- 4. Separate infra from growth: while your team builds on AgentCore, consider letting Planetary Labour run distribution and SEO on autopilot.
Questions about MCP and agent tooling more broadly? See our MCP security guide for how to keep agent-to-tool connections safe in production.