What does a real production agent look like under the hood? 5 agents on Astropods
Originally published on the Postman blog.
What does a real production agent look like under the hood? Most agent demos never tell you. They end the moment the agent answers a question. You never see how it was wired up, what it talks to, or whether it would hold up in production.
If you're building or evaluating agents, that hidden part is what decides whether yours survives contact with real users. This post opens up five real agents running on Astropods. For each one you get the config, the interface choices, the human-in-the-loop decisions and the patterns you can borrow. Several of them are public blueprints you can deploy yourself with your own keys.
Agents in Action
Agents in Action is a video series built around that missing part. Each episode opens up a real agent running on Astropods: the running system, the config behind it, and a short conversation with the person who built it about the choices they made. No slides, no toy demos.
Astropods in 60 seconds
Astropods is the agent operating system layer in the Postman AI stack. It runs and controls agents in production, with runtime, observability and guardrails built in.
Every agent below follows the same lifecycle:
- Project: your code plus an
astropods.ymlspec for models, interfaces, integrations, knowledge stores and runtime needs. - Blueprint: the project published as a versioned blueprint.
- Deployment: the blueprint deployed as a running agent.
- Observability: the running agent watched through traces, evals, token usage, spend and latency.
ast project create my-agent ast deploy The same spec drives local dev and production, so what you see in these episodes is what you can run yourself.
1. The Reading Room: an agent with its own UI
This agent tackles newsletter overload with an interface that isn't a chat box. It has a library entrance and digital bookshelves, and it renders newsletters as books you read chapter by chapter.
Under the hood
- Pulls newsletters from Gmail using a label and an app password.
- Uses Anthropic Claude to extract key insights, highlights and full article text.
- Stores everything in a serverless Neon database.
- Astropods hosts the custom web frontend and auto-provisions pods, domains and hosting.
- Digests run manually or on a schedule.
Steal this: if your users need to browse, compare or reread what an agent produced, give it a purpose-built UI. The Reading Room is available as a blueprint.
2. Ada: a coding agent that lives in the cloud
Ada is a cloud-native autonomous coding agent built directly on Astropods. It isn't trapped on your laptop or locked to one model provider.
Under the hood
- Runs in a sandboxed terminal in the cloud.
- Two ways to talk to it: a custom terminal frontend built on AG-UI, or the built-in Astropods Chat playground.
- The Astropods AI Gateway routes models, so you can swap between Claude, Kimi K3 and self-hosted models, including open-weights and on-prem options.
- Persistent memory lives in PostgreSQL, with access controls and identity management.
- Deploys from the Blueprint Registry, with automated GitHub CI/CD syncing.
- The team used Ada to build and iterate on Ada itself.
Use this: route models through a gateway so that swapping providers is a config change, not a rewrite. Keep an agent's context somewhere the whole team can benefit from it.
3. Memory Box: search everything you saved
That X bookmark you can't find. The GitHub star you saved months ago and can't name. Memory Box, built by a community member, consolidates your saved items into one place and lets you find them by asking.
Under the hood
- Vectorization powers natural-language search across your saved content.
- A chat interface handles retrieval.
- Every deployment is sandboxed.
- Setup needs an Anthropic key and a few steps.
Use this: a single-purpose agent with a narrow job is easy to build, easy to reason about and easy to deploy.
4. Cruise Line: code review with you in the loop
AI is generating more code, PRs are piling up, and many review tools respond by flooding them with noisy comments. Cruise Line flips that: AI analyzes the PR, and you decide which comments are worth posting to GitHub. Nothing posts until a human approves it.
Under the hood
- Automated analysis covers security, performance, code reuse and maintainability.
- Custom rule sets enforce team-specific standards, such as blocking plain string SQL injections.
- A built-in chat lets you ask questions about the code without pulling it down locally.
- You deploy it by connecting your GitHub repo and adding an Anthropic key.
- It's published as a public blueprint, so any team can deploy its own instance.
Use this: keep a human in the loop for any action that's visible to others. The agent proposes, you approve.
5. Industry News agent: a daily briefing in Slack
This one is a good template for deploying any agent to Slack. A public blueprint gives anyone on your team a newsletter-style briefing from a single @mention.
Under the hood
- Queries NewsAPI, GNews, The Guardian and MediaStack in parallel, then removes duplicate articles.
- You set the required and optional API keys in the deployment form.
- Slack setup needs:
- an app-level token with
connections:write - Socket Mode
- Event Subscriptions with
app_mentions:read - bot token scopes for channel history, posting and direct messages
- an app-level token with
Use this: fan out to several sources at once and dedupe before the model sees anything. Then put the agent where your team already works.
Patterns that show up across the series
- Narrow beats broad. Memory Box, Cruise Line and the news agent each do one job well.
- The interface is part of the agent. Chat, a custom UI, a cloud terminal and Slack each fit a different job.
- Don't lock yourself in. Ada routes models through the AI Gateway, so swapping providers is a config change, not a rewrite.
- Keep humans in charge of visible actions. Cruise Line lets you approve before anything posts.
- Ship it, then watch it. Traces, evals and spend monitoring come with the platform rather than being bolted on later.
Quick-start: run one yourself
The original post lays out a 10-minute quick-start path:
- Install the
astCLI and start Docker. - Browse the Blueprint Registry at astropods.com, pick an agent from the series (The Reading Room, Ada, Cruise Line or the Industry News agent) and deploy it with your own keys. Or start your own project:
ast project create my-agent ast deploy - Connect it to Slack with
--adapter slack, then check its traces and token usage once it's live.
Watch the series and go deeper
The full post on the Postman blog links each blueprint, the Agents in Action playlist, the Astropods docs and the community Discord. Watch an episode, then deploy one of the agents from the Blueprint Registry with your own keys.
Which pattern would change how you're building agents today: narrow scope, a model gateway, human approval or built-in observability? Tell me in the comments.