And with that... ๐๐ด๐ฒ๐ป๐๐๐ผ๐ฟ๐ฒ ๐ต๐ฎ๐ฟ๐ป๐ฒ๐๐ ๐ถ๐ ๐ป๐ผ๐ ๐๐! ๐
Every AI agent needs the same scaffolding underneath: tools to interact with, an environment to work in, a system to manage context, memory to personalize, identity to control access, and observability to understand and course correct. Wrapped in a loop that calls the model, picks a tool, and recovers from failures. This is the agent harness, and until we launched it in preview, every team using AgentCore built it from scratch.
Today, the managed agent harness in AgentCore is generally available. You declare your agent and run it in two API calls. You point to the model, tools, skills, and instructions as configuration, and AgentCore stitches together everything around it to make the agent production-ready. What you get out of the box:
1๏ธโฃ ๐๐ป๐ ๐บ๐ผ๐ฑ๐ฒ๐น, ๐๐๐ถ๐๐ฐ๐ต ๐บ๐ถ๐ฑ-๐๐ฒ๐๐๐ถ๐ผ๐ป. Bedrock, OpenAI, Gemini, or any LiteLLM-supported provider. Switch providers mid-session without losing context. Plan on one model, code on another, summarize on a third.
2๏ธโฃ ๐ง๐ผ๐ผ๐น๐, ๐ฑ๐ฒ๐ฐ๐น๐ฎ๐ฟ๐ฎ๐๐ถ๐๐ฒ๐น๐. MCP servers, AgentCore Gateway, built-in Browser and Code Interpreter, or your own inline functions. One config line per tool, with no boilerplate to write.
3๏ธโฃ ๐๐๐ถ๐น๐-๐ถ๐ป ๐บ๐ฒ๐บ๐ผ๐ฟ๐, ๐๐๐ฎ๐๐ฒ๐ณ๐๐น ๐ฏ๐ ๐ฑ๐ฒ๐ณ๐ฎ๐๐น๐. Memory is on out of the box, persisting users and conversations across sessions. Each session also runs in its own secure, isolated microVM with a filesystem and shell.
4๏ธโฃ ๐ฆ๐ธ๐ถ๐น๐น๐ ๐ผ๐ป ๐ฑ๐ฒ๐บ๐ฎ๐ป๐ฑ. Compose your agent with bundles of markdown and scripts that give it domain knowledge. Pull from git, S3, your own path, or flip on the full AWS-curated catalog with a single ๐๐ ๐๐๐๐๐๐๐ toggle. The harness loads and executes them only when the task calls for it.
5๏ธโฃ ๐ฅ๐๐ป ๐ถ๐ป ๐๐ต๐ฒ ๐ฒ๐ป๐๐ถ๐ฟ๐ผ๐ป๐บ๐ฒ๐ป๐ ๐๐ผ๐ ๐ป๐ฒ๐ฒ๐ฑ. Use the default image, or bring your own container with your source, runtimes, and dependencies. Persist files with managed session storage, or mount EFS or S3 Files for state that outlives a session.
6๏ธโฃ ๐ข๐ฏ๐๐ฒ๐ฟ๐๐ฒ, ๐ฒ๐๐ฎ๐น๐๐ฎ๐๐ฒ, ๐ผ๐ฝ๐๐ถ๐บ๐ถ๐๐ฒ. Unified observability gives you one view across every primitive the agent touched. Score real traffic with built-in evaluators, get prompt and tool recommendations, and A/B test them on live traffic before you roll out.
7๏ธโฃ ๐ฉ๐ฒ๐ฟ๐๐ถ๐ผ๐ป, ๐ฟ๐ผ๐น๐น ๐ฏ๐ฎ๐ฐ๐ธ, ๐ฎ๐ป๐ฑ ๐ฒ๐ ๐ฝ๐ผ๐ฟ๐. Every update is an immutable version, so rollback is pointing an endpoint at an earlier one. And when config stops being enough, export to Strands-based code and keep running on the same compute, same microVM, same observability. No re-architecture, no platform tax.
Trying a new model or tool is a config change, not a code rewrite. Managing context, remembering across users, adding a skill, tuning a prompt: again config, not infrastructure. The whole loop (build, connect, optimize, ship) collapses from weeks into minutes.
Read more in our launch blog: https://t.co/seJp4flUOh
This article reminds me of why I voted for Hillary back in 2016. A weak candidate in retrospect, she genuinely cares about America and its social fabric, unlike her opponents who prosper from its disintegration.
https://t.co/ucFFyy32bm
Interesting article by @jon_rauch about the need for structure and the limits of self-organization in politics. "The world of unstructured interpersonal exchange looks a lot like Twitter," he writes, "an epistemic version of Hobbes's state of nature."
https://t.co/g0sPm3px2L
@kohidave You also make your app less portable, and now to actually develop on it, you have to deploy to a dev environment on AWSโyou can't just run it locally anymore in a nice little HTTP server that does it all.
5/5
@kohidave You can use AWS solutions to solve these problems, but then it becomes harder to tell what's going on, since you end up moving business logic into configuration, e.g., you move routing into an API Gateway config and now your actual app code doesn't do that anymore.
4/n
@kohidave The other option is to write code for Lambda from the start, but then you have to solve a bunch of problems on your ownโrouting, auth, etc.
3/n
@kohidave To deploy these kind of apps to Lambda, you need to, first, know to use API Gateway to make the Lambda function callable via HTTP, and, second, you need to write some code to adapt your existing code to the Lambda handler signature, which is kind of a pain.
2/n
@kohidave I have a take @kohidave! Many of the major application frameworks that many companies use are based on HTTPโRails, Django, Express, Spring, etc. And they all have a nice dev experience where you run a little server locally. And great docs and tutorials made by the community.
1/n
@nickfrenchnyc Ah, being out here in Seattle, I don't have my finger on the pulse of NYC politics. But I liked the idea of a philosopher like you in public office. ๐
@DrTexx @jeremy_soller @andrewbadr was working on a similar project called @MixilyEvents, but it looks like its focus has changed to online events. I'd be interested to know if you find a solution ๐