Bedrock AgentCore Managed Harness Reaches GA in 2026
AWS Bedrock AgentCore's Managed Harness is now GA — build production AI agents with just a model, prompt, and tools. See what's new and try it today.
Lead Data Scientist
If you've built an agent on AWS in the last year, you know the drill: pick a model, wire up an orchestration loop, bolt on memory, figure out tool-calling, containerize it, and then start worrying about evaluation and safety. AgentCore has been chipping away at that stack piece by piece since its 2025 launch — but the last few months are the first time it's felt like the pieces actually snapped together into something you can hand to a junior engineer on day one.
Here's a tour of what shipped, what it replaces in a typical build, and a quick example to try today.
The headline: Managed Harness is now GA
Instead of writing your own agent loop — reasoning → tool selection → action → response — the Managed Harness lets you define an agent with three inputs: a model, a system prompt, and a set of tools. AgentCore runs the loop for you, in an isolated microVM per session with its own filesystem and shell.
What makes this more than a toy quick-start:
- It's model-agnostic. Swap models mid-session, or override any config at invocation time without redeploying.
- It doesn't lock you in. Once your prototype outgrows the defaults, you can export the harness's orchestration into Strands-based code and keep going from there — no rebuild required.
- It ships with memory, skills, and web browsing by default, not as separate integrations you have to stitch on.
A minimal definition looks roughly like this:
Python
import boto3client = boto3.client("bedrock-agentcore")response = client.create_harness(name="support-triage-agent",model="anthropic.claude-sonnet-5",system_prompt="You triage incoming support tickets and draft a first response.",tools=["web_search", "knowledge_base_lookup"],)
That's genuinely the whole "getting started" step. Everything else — scaling, session isolation, observability — is inherited from the platform.
Agents can now ground themselves in the live web, without leaving your VPC boundary
A new managed Web Search tool exposes a built-in connector on the AgentCore Gateway over MCP, so an agent can pull current information without you standing up your own search integration or sending queries to a third-party vendor. It returns ranked snippets with sources, titles, and publication dates, and — notably — keeps the query traffic inside your AWS security boundary rather than round-tripping to an external API.
Paired with the new Fully Managed Knowledge Base (native connectors for S3, SharePoint, Confluence, Google Drive, plus an agentic retriever for multi-step queries), this closes a gap a lot of teams were hand-rolling with custom RAG pipelines: your agent can now reach organizational knowledge and the live web through the same gateway, with one governance model instead of two.
The part nobody talks about enough: agents that improve themselves
This is arguably the more interesting shift. AgentCore now closes the observe → evaluate → improve loop natively:
- Recommendations analyze production traces and evaluation results, then suggest concrete edits to your system prompt or tool descriptions — grounded in how the agent actually behaved, not guesswork.
- Batch evaluations score a proposed change against your test set before it goes anywhere near production.
- A/B testing splits live traffic between agent versions and reports statistical significance before you commit.
Every suggested change still requires your sign-off — this isn't autonomous self-modification — but it turns "why did quality quietly degrade over the last month" from an investigation into a dashboard. And it works whether your agent runs on AgentCore Runtime, Lambda, EKS, or somewhere entirely off-AWS.
Headroom, if you're scaling
As of July 2026, default runtime quotas jumped meaningfully: up to 5,000 concurrent sessions in us-east-1/us-west-2 (2,500 elsewhere), 200 invocations/second per agent, and faster new-session creation for container deployments. If a quota limit was the reason a proof-of-concept never got a production date, it's worth re-checking the numbers — these are automatic, no request needed.
Try this today
If you've been putting off a real agent build because the orchestration boilerplate felt like a week of throat-clearing before the interesting part, the Managed Harness is the fastest way to find out if AgentCore fits your use case:
- Define a harness with one tool and a narrow system prompt.
- Point it at Web Search or a small knowledge base.
- Run it, look at the trace, and see what a Recommendation suggests after a day of real traffic.
That loop — build, observe, improve — used to take a season. Now it's closer to an afternoon.
What are you building with AgentCore right now? Drop it in the comments — always looking for real workloads to compare notes on.
#aws #bedrock #agentcore #ai #machinelearning
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article