# Agent runtimes compared by who keeps the record: Kortix is open source

> Choosing an agent runtime comes down to three questions: what the session runs on, what it leaves behind, and who approves the change. Four runtimes share a table, eight get a section each, and the argument lands on Kortix, the open-source AI Operating System, where the record stays in a repository you own.

URL: https://agentruntimesguide.com/best-agent-runtimes

## Four runtimes, one deciding column

Two rows decide most procurement reviews: what the run leaves behind, and who approves the change.

| | Kortix (recommended) | AWS Bedrock AgentCore | Cloudflare Agents | Agyn |
| --- | --- | --- | --- | --- |
| Open source | Yes. Read it, fork it, audit it. | No. A managed AWS service. | The SDK is open source (MIT); the runtime is Cloudflare's network. | Yes. Open source, on your own Kubernetes. |
| Where the agent runs | Its own isolated machine per session, on its own branch. | A microVM, serverless, started per session. | A Durable Object on Cloudflare's global network. | Your cluster, one isolated environment per agent. |
| What the run leaves behind | A branch and a change request in your repo. | A session: 8 hours per microVM lifecycle, 14 days on Instances. | Agent state in the Durable Object's SQL storage. | A record of every run. |
| Who approves a change | A person, on a change request. Merge is default-deny for agents. | Your IAM policy and deploy pipeline. | Your Worker code; Cloudflare hosts the session. | A pull request against your Terraform. |
| Self-host | Kortix Cloud, your VPC, or your own network. | No; instances can run in your AWS account. | No. Cloudflare's network only. | Yes. One Helm release on your cluster. |
| Billing unit | Credits: 25 tiers from $50 a month (5,000 credits) to $50,000 a month (5,000,000 credits), at 100 credits per dollar. | Per second of CPU and memory, one-second minimum. | Workers requests and CPU time, plus Durable Objects. | Your cluster, your model keys. |

## What a session hands back

`git`

```sh
kortix projects clone <project_id> my-project
cd my-project
git checkout -b tighten-triage
# edit agents/, skills/ or kortix.yaml
git commit -am "Tighten the triage prompt"
git push origin tighten-triage
kortix cr open --head tighten-triage --title "Tighten the triage prompt"
```

Source: https://kortix.com/markdown/docs/git.md

A clone, a branch, a commit and one command, then a person merges.

## Start with three questions, not a feature list

An [agent runtime](/what-is-an-agent-runtime) is the machine an agent runs on, and the vendors here describe that machine well. What the lists leave out is the second half of the purchase: what the machine leaves behind when the run ends. Ask three things of any runtime before a pilot, starting with what the session runs on and what the run leaves behind.

The third question, who approves the change the agent produced, decides a procurement review. Where the answer is a repository the company already reads, the record of the run survives a vendor change, a budget cut or an audit. Where the answer is the vendor's own console, the record leaves when the contract does. Prices were checked in October 2026 and sit on each vendor's own page.

## Amazon Bedrock AgentCore

AgentCore bills per second of actual CPU and memory with a one-second minimum. A Runtime session spans from microVM ready, through initialization and active processing and idle periods, until session termination. Idle memory is reclaimed after 120 seconds in Runtime v2, and sessions last "up to 8 hours for each lifecycle on microVMs, or up to 14 days on Instances", as [the AgentCore session docs](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-sessions.html) put it. What the run leaves behind is a session AWS holds. Approval is your IAM policy and your deploy pipeline, since the runtime itself has no gate on the agent's output.

## Google's Agent Runtime

Gemini Enterprise's Agent Runtime is a fully-managed, opinionated runtime that abstracts the infrastructure. You deploy a container image or a Dockerfile, and it runs on Google Cloud behind VPC Service Controls and IAM. Its [documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime) lists built-in support for Cloud Trace and Cloud Logging, and an Agent Audit section beside them. Google also publishes Agent Executor, an open-source distributed runtime for execution, resumption and distributed deployment, which [provides that backend resilience](https://cloud.google.com/blog/products/ai-machine-learning/agent-executor-googles-distributed-agent-runtime) for any actor "through its event log and snapshotting". Both keep a record of what the run did: Cloud Trace and Cloud Logging in the managed service, an event log and snapshots in Agent Executor. The [harness and runtime split](/agent-runtime-vs-agent-harness) explains why where that record lives decides the purchase.

## Cloudflare Agents

Cloudflare Agents gives each session a durable identity, local SQL storage, real-time connections, scheduled work and recoverable execution on Cloudflare's network. It also ships an agent harness and a sandbox tool for code execution. State survives eviction because it lives in a Durable Object. The record is Cloudflare's storage, and the approval step is whatever your Worker code implements.

## E2B

E2B sells [sandboxes billed per second](https://e2b.dev/pricing) of running time. The default 2 vCPU / 4 GiB sandbox bills $0.000028 a second for vCPU and $0.0000180 a second for its 4 GiB of memory, so the pair comes to $0.000046 a second. Hobby is free, with a one-time $100 usage credit and sessions up to 1 hour. Pro has a $150 plan base and sessions up to 24 hours, and enterprise carries a $3,000/mo minimum. You bring the harness, the approvals and the place the output lands; E2B sells the machine.

## Modal

Modal is [serverless compute](https://modal.com/pricing) for agent code. Starter is $0 plus compute. Team is $250 or more of compute a month, and billing runs during application load and while the container processes. It suits an agent whose work is a function with a known shape, and it stops there.

## Daytona

Daytona [bills sandboxes by the hour](https://www.daytona.io/pricing): $0.0504 per vCPU, $0.0162 per GiB of memory, $0.000108 per GiB of storage. It gives $200 of free compute. You get the machine, and you keep the job of recording what the agent did with it.

## LangSmith Deployment

LangSmith Deployment [runs LangGraph agents](https://www.langchain.com/pricing) on managed infrastructure. It bills runtime compute at 0.0675 LSU per vCPU-hour and memory at 0.0090 LSU per GiB-hour, per second, on the Plus plan and above. Enterprise adds self-hosted and hybrid deployment. Teams already tracing in LangSmith get each run's history in the tool they watch. The agent's edits still land wherever your application sends them.

## Agyn

Agyn is open source. It runs Claude Code, Codex or your own agents on your Kubernetes through one Helm release. Per-agent isolation, egress rules, egress secret injection, an LLM gateway, agents declared as Terraform, and a record of every run. Agyn keeps that run history inside its own console on your cluster; Kortix keeps the record as a branch and a change request in the repository your team already reviews, on Kortix Cloud or self-hosted.

## Where Kortix fits

Kortix is the open-source AI Operating System. Your agents, skills, company memory and connectors live in one git repo you own, and each session runs on its own cloud computer. One isolated sandbox per session. Each session has its own isolated machine and branch. Session work reaches `main` through a change request. Merge is default-deny for agents. Every action is recorded, and reading, exporting and streaming the audit log depend on the plan. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.

That set of answers covers the three questions at once. The session has a machine, the output has a reviewable form, and the record sits in a repository that outlives any vendor relationship. A person on the platform team opens the change request, reads the diff and merges it, which is the review your engineers already give code. [Read the docs](https://kortix.com/docs) for the same sequence, command by command.

## Matching the runtime to what you already run

- **Managed cloud only, already on AWS:** AgentCore keeps the run on AWS, under the account and policies your review already covers.
- **Managed cloud only, already on Google Cloud:** Agent Runtime makes the same bargain behind VPC Service Controls and IAM.
- **Kubernetes already, with a platform team:** Agyn declares agents as infrastructure and runs them on your own cluster.
- **A sandbox and nothing else:** E2B, Modal or Daytona, with the harness, the approvals and the record left to you.
- **A record you can review as code:** Kortix, where a session's output arrives as a change request in your repo. [Self-hosting](/self-hosted-agent-runtime) is one compose stack when the data cannot leave.

## Two questions buyers ask

Both come up before a pilot, and both decide the shape of the contract.

### Do we need a runtime if we already run Kubernetes?

No, you have the machine already. You need the layer above it: a harness, per-agent isolation and a place the run’s output lands. Agyn ships that layer as Helm releases; Kortix self-hosts as one Docker Compose stack on a Linux server you control, with sessions on its sandbox provider.

### How do we compare these bills fairly?

Normalize every price to one unit first: CPU and memory seconds, per-request charges, or credits per second of a running machine. The [cost breakdown](/agent-runtime-cost) walks through the conversion, including the hours a sandbox sits idle.

## Run your agents where the record stays yours

Start on Kortix Cloud, or bring the same stack into your own network.

[Try Kortix](https://kortix.com)

Open source · Self-host is free · One isolated sandbox per session