AWS Bedrock AgentCore is managed infrastructure for running AI agents — the compute, memory, identity, and tool-access layer that agents execute on within AWS. It uses Cedar for fine-grained authorization, the same policy language Olivares AI uses for its governance engine. The convergence on Cedar is the interesting part; the divergence is the layer each product occupies.
Note: This comparison covers AgentCore as the agent runtime infrastructure layer. AWS Bedrock also includes model access, Guardrails (content safety), and Knowledge Bases — those adjacent capabilities are covered in vs AI gateways & Guardrails.
What AgentCore does well (use it for this)
AgentCore is AWS’s answer to the operational problem of running agents at scale on managed infrastructure:
- Agent runtime. Managed compute for agent execution — provisioning, scaling, and lifecycle without operating your own infrastructure.
- Tool access. Managed connectivity between agents and AWS services (databases, APIs, storage) with built-in tool integrations.
- Cedar authorization. Fine-grained, attribute-based access control using Cedar — the same policy language, the same formal verification properties.
- Observability. Built-in monitoring, logging, and tracing for agent execution within the AWS ecosystem.
- Identity. Agent identity tied to IAM — agents as principals within the AWS identity model.
If your problem is “run agents on AWS with managed infrastructure, IAM-scoped tool access, and Cedar authorization,” AgentCore solves it. We do not reimplement that runtime layer.
The governance gap above the runtime
AgentCore governs what agents can do within the AWS runtime boundary. It does not answer the questions that arise when agents operate across infrastructure boundaries — multiple clouds, on-premises databases, MCP servers, self-hosted tools — or when the evidence needs to survive independently of the platform that produced it.
| Gap the runtime leaves | Why it matters | What Olivares AI provides |
|---|---|---|
| Cross-infrastructure scope | AgentCore governs within AWS; many estates span AWS, Azure, on-prem, and self-hosted tooling | Infrastructure-wide governance across any environment — cloud, on-prem, air-gapped — through one access map |
| Tamper-evident evidence | Runtime logs are mutable platform records; an auditor wants independently verifiable proof | Append-only, hash-chained, Ed25519-signed ledger — off-box verifiable, exportable as OSCAL evidence |
| Vendor-neutral deployment | AgentCore requires AWS; the governance data lives in AWS’s managed service | Self-hosted on your infrastructure — Linux, Docker, Kubernetes, air-gapped; the governance data never leaves your perimeter |
| Framework-neutral governance | AgentCore governs agents running on its runtime; agents using other runtimes need separate governance | Governs agents from any framework or runtime — Claude Code, AutoGen, LangGraph, custom — through one plane |
| Live-session intervention | Runtime-level controls; a kill switch or break-glass across live sessions is not documented | HITL approvals, break-glass, and a kill switch that denies all governed actuation until dual-control re-enable |
| Sovereignty | Managed service on AWS — the governance data lives in AWS’s cloud | Self-hosted; the access relationships and the governance record stay on your infrastructure |
These are not runtime features. The gap is governance and evidence that sits above any single runtime, not a better runtime.
The Cedar convergence
The most significant shared design decision: both Olivares AI and AWS Bedrock AgentCore use Cedar for authorization. This is not a superficial overlap:
- Same policy language. Cedar’s deny-by-default, attribute-based model provides formal verification properties that ad-hoc RBAC cannot match.
- Same mental model. An organisation that writes Cedar policies for AgentCore’s runtime can write Cedar policies for Olivares’ governance plane with the same syntax, the same semantics, the same tooling.
- Composable layers. AgentCore enforces Cedar at the runtime tool-access level; Olivares enforces Cedar at the infrastructure governance level. Same language, different scope, same policy store if you choose.
The convergence on Cedar is the reason these two products compose cleanly rather than conflicting.
When AgentCore is the right choice
- You are building on AWS and want managed agent runtime infrastructure — provisioning, scaling, tool connectivity, IAM-scoped identity — without operating your own compute layer.
- Your agent estate is AWS-native and the governance boundary matches the AWS account boundary.
- You need Cedar authorization at the runtime level and AWS’s managed observability is sufficient for your compliance requirements.
When Olivares is the right choice
- Your agent estate spans multiple clouds, on-premises, and self-hosted infrastructure — and you need one governance plane across all of it.
- The evidence needs to be tamper-evident, independently verifiable, and exportable — not tied to any single platform’s logging infrastructure.
- You need a self-hosted or air-gapped control plane where the governance data never leaves your infrastructure perimeter.
- You need vendor-neutral governance that works the same way whether agents run on AgentCore, on bare metal, or in your own Kubernetes cluster.
When they compose
The strongest deployment pairs both:
- AgentCore as the managed runtime — runs your agents on AWS, provides IAM-scoped tool access and Cedar authorization at the runtime boundary.
- Olivares as the governance plane — provides infrastructure-wide visibility, the cross-runtime access map, tamper-evident evidence, and live-session intervention (kill switch, break-glass, HITL) across agents on AgentCore and agents elsewhere.
Same Cedar language at both layers. One policy mental model. The runtime governs what the agent can call; Olivares governs what it actually reached, and proves it.
Related
- vs AI gateways & Guardrails — Bedrock Guardrails as a content-safety hook, not a competitor.
- vs AI control towers — where the tower fits above the runtime and the governance plane.
- Governing subscription-authed agents — how Olivares governs agents that authenticate with a user subscription.