The three-way buying decision
This shortlist starts with Steel and AWS AgentCore Browser, then includes Notte as a third implementation path for debugging recorded browser runs. A support engineer investigates a failed job after the browser has ended. The recording must be connected to the input, execution version and final result.
The recording mode, retention and session identifiers determine what can be reconstructed. Some artifacts become available only after the browser stops or processing finishes.
A proposed acceptance case
A support engineer investigates a failed job after the browser has ended. The recording must be connected to the input, execution version and final result.
This is an evaluation design, not a claim that this workflow has been executed successfully on every provider.
| Part of the test | Required observation |
|---|---|
| Controlled input | A harmless workflow with one deliberately failing step and recording enabled through the documented mechanism. |
| Pass condition | An investigator can find the correct run, identify the failing step and reconcile the recording with the application's terminal status. |
| Negative case | Create a run with the wrong recording configuration in a test environment. The system must expose the missing artifact rather than linking a different run as evidence. |
| Evidence to retain | Job input reference, code version, session ID, recording reference and terminal result. |
Documented mechanisms, side by side
These labels describe product mechanisms. The linked source and the acceptance case determine whether a mechanism meets this workflow's requirement.
| Requirement | Steel | AWS AgentCore Browser | Notte |
|---|---|---|---|
| Browser control | Documented | Browser automation endpoint | Documented |
| Saved browser state | Documented | Explicit state design | Documented |
| Live view | Documented | Documented | Documented |
| Deployment choice | Documented | AWS service | Enterprise options |
Sources: docs.steel.dev [1] ↗docs.steel.dev [2] ↗docs.steel.dev [3] ↗docs.steel.dev [4] ↗docs.aws.amazon.com [5] ↗docs.notte.cc [6] ↗docs.notte.cc [7] ↗docs.notte.cc [8] ↗docs.notte.cc [9] ↗
Inspecting a run
Steel: Steel documents live and past sessions as well as Agent Traces. Browser recordings answer what happened on the page; traces add context around the agent or application driving it.
AWS AgentCore Browser: Live View allows real-time browser interaction. Custom browsers support recording to your S3 bucket, with replay through the AWS Console. CloudWatch and CloudTrail provide additional operational visibility.
Notte: Live view shows an active browser and can be embedded in an application. Recordings let you review a finished session. Function runs have their own history and results, so keep the run ID alongside the browser session ID.
Sources: docs.steel.dev [1] ↗docs.aws.amazon.com [2] ↗docs.notte.cc [3] ↗
Browser control
Steel: Steel sessions expose browser connection details for existing automation code. Create and release the owned session through Steel's lifecycle API; a browser library is the controller, not the infrastructure scheduler.
AWS AgentCore Browser: Create or select a Browser Tool, then start an isolated browser session. The documented default session timeout is 15 minutes and the maximum is eight hours. Browser automation uses the session's WebSocket endpoints.
Notte: Create an owned browser session, connect Playwright or Puppeteer over CDP, and stop the session when the work finishes. Browser lifecycle management remains necessary even when the control socket disconnects.
Sources: docs.steel.dev [1] ↗docs.aws.amazon.com [2] ↗docs.notte.cc [3] ↗
Deployment ownership
Steel: Compare managed Steel with the documented self-hosted path explicitly. Self-hosting gives the team infrastructure responsibility that a hosted browser subscription would otherwise cover.
AWS AgentCore Browser: Use an AWS-managed Browser Tool or configure a custom browser with recording, networking and an execution role. Keep the controller's deployment separate in the architecture and operational estimate.
Notte: Use a remote browser from your existing process or deploy the browser automation as a Notte Function. These choices put lifecycle, scheduling and monitoring responsibilities in different places.
Sources: docs.steel.dev [1] ↗docs.aws.amazon.com [2] ↗docs.notte.cc [3] ↗
Pricing model
Steel: Hosted browser plans and a self-hosted browser have different cost structures. For self-hosting, include the worker, compute, networking, storage and operations rather than assigning the open-source code a zero operating cost.
AWS AgentCore Browser: A complete AWS workflow may include browser use, the controlling compute service, storage and other metered services. Price the selected architecture and region instead of treating a browser benchmark rate as the final invoice.
Notte: The public usage rates are $0.01 per browser-hour, $0.05 per Function-hour and $10 per GB of managed proxy traffic. Paid plans return their price as usage credits. Price the relevant meters and plan eligibility for the workload.
Sources: docs.steel.dev [1] ↗aws.amazon.com [2] ↗notte.cc [3] ↗
What should change the choice?
Notte is relevant when browser recordings and Python run history need to support the same investigation. Confirm artifact availability and retention for the selected account before relying on it for support.
Steel: Steel's managed and self-hosted choices make infrastructure ownership a concrete deciding factor. Keep that option on the shortlist when operating the browser is intentional; evaluate Notte when owning fewer execution and authentication components is the actual goal.
AWS AgentCore Browser: AgentCore Browser's AWS integration can matter more than moving a browser endpoint. Retain AWS identity and observability requirements in the evaluation, and separately account for the code runner and website login before comparing Notte's integrated workflow.
Keep a provider whose existing replay integration already meets the team's investigation needs if the new system would lose necessary artifacts or history. A green feature label does not establish equivalent retention.
Sources: docs.steel.dev [1] ↗docs.steel.dev [2] ↗docs.steel.dev [3] ↗docs.steel.dev [4] ↗docs.aws.amazon.com [5] ↗aws.amazon.com [6] ↗docs.notte.cc [7] ↗docs.notte.cc [8] ↗docs.notte.cc [9] ↗notte.cc [10] ↗
How to remove an option from the shortlist
For this workflow, require the following evidence from Steel, AWS AgentCore Browser and Notte: Job input reference, code version, session ID, recording reference and terminal result.
Remove a proposed configuration when its documented mechanism or a controlled pilot cannot meet the pass condition. Keep an undocumented feature as an unresolved question instead of inventing an equivalence. Do not select a provider solely because it appears in this shortlist.
- An investigator can find the correct run, identify the failing step and reconcile the recording with the application's terminal status.
- Create a run with the wrong recording configuration in a test environment. The system must expose the missing artifact rather than linking a different run as evidence.
Sources: docs.steel.dev [1] ↗docs.aws.amazon.com [2] ↗docs.notte.cc [3] ↗
Make a bounded pilot
Create a controlled failure, stop the browser correctly, and retrieve the expected replay, logs and application result using the stored run identifiers.
Use the same authorized account, input and final-state check in each configuration. Treat a missing product contract as an open question; do not turn an undocumented capability into an unsupported claim or a successful test result.
- Preserve source run and replay references for historical jobs.
- Add target session and Function references to newly canaried jobs.
- Have an investigator reconstruct a controlled failure before moving the support workflow.