In partnership with
Anthropic shipped Claude Managed Agents back in April and Google followed suit in the I/O last week.
Most of the coverage in this context has been about enterprise orchestration, multi-agent fleets, governance, etc.
But hardly anyone's talking about how useful it is on an individual level too.
Every developer I know has the same guilty tab open, whether it's a Reddit saved search you forgot about, or the HN thread someone shared that you bookmarked and never opened.
The information you care about lives across half a dozen surfaces, and the synthesis happens in your head, when it happens at all.
And I think Managed Agents can help here. (among many other indie dev workflows)
So we'll be building a personal information agent that wakes up every morning, searches three topics you have asked for, and emails you a tight briefing at 7 AM.
Before we dive in, a quick poll:
How do you currently keep up with topics you care about professionally?
Before We Begin
What Managed Agents Gives You
6 months ago, building something like this meant stitching together your own container runtime, state management, credential storage, error recovery, and a sandbox so your agent couldn't wipe your laptop by mistake.
Managed Agents covers all of that in three API calls.
You define the agent itself (model, system prompt, and tools available to it), point it at an environment that holds the container config, and start a session each time you want it to run.
The toolkit the agent has access to via agent_toolset_20260401 includes bash, file operations, web search, web fetch, grep, and glob.
Enough for it to read from the open web and write output to disk, while API keys stay in the credential vault, encrypted on storage and pulled at runtime, so they don't end up in logs.
Standard Claude API token rates apply, with an extra $0.08 per session-hour measured in milliseconds.
A 5-minute run costs a few cents of compute and maybe $0.01–0.03 in tokens.
Means, three topics, daily, for a full month works out to roughly $1–3. That’s cheaper than the Substack you stopped reading.
The Architecture
The agent definition lives permanently on Anthropic's side once you create it.
Your scheduler (system cron, or whichever box runs your jobs) kicks off a new session each morning, and once the session finishes, an email step pushes the briefing to your inbox.
The flow:
Sessions don't carry state between runs.
Each morning the agent starts cold, pulls the latest on your topics from the open web, opens yesterday's briefing (stored as JSON), and writes a new one covering what's changed.
The state file is the only memory across runs, so no database or Redis to maintain.
The Code
So you'll need the Anthropic Python SDK, a cron scheduler (system cron is fine), and an email sender (Resend, Postmark, plain SMTP) you already trust.
Step 1: Create the agent (run once)
Run once, save the agent ID that comes back.
The ID is permanent and every future session you spin up references it.
Step 2: Create the environment (run once)
Networking is unrestricted because the agent needs to search and fetch from the open web.
If you're on a self-hosted sandbox (public beta as of May 19), you can lock this down further with your own policies.
Step 3: The daily runner (runs on cron)
The extract_briefing function is about ten lines of parsing the session output for the final markdown.
I've left it out since your email setup will dictate the format.
Step 4: The cron entry
Runs at 7 AM local time, every day.
If you'd rather use UTC (the Managed Agents default), adjust the offset.
What You Should Track
I narrowed down to just three topics after a few iterations, because two felt thin in practice and five became too much for a morning read.
Three keeps the briefing scannable while still leaving room for cross-domain stuff to come through.
You can swap topics whenever.
The file is read fresh on every run, so updating today means tomorrow's briefing reflects it.
Where This Breaks
The agent only sees the open web.
Paywalled articles, private Slack channels, and internal Notion docs stay invisible unless you add an MCP server for the source.
Production MCP servers exist for Slack, Notion, and GitHub among others, but without one, the briefing covers public signal only.
The sessions take 10–20 seconds to start, and a full three-topic run usually lands at 2–4 minutes.
If your cron timeout is tight, give it room.
Cold starts mean the agent has no memory between runs except what's been persisted to disk.
And cost grows with topics and depth.
Three topics on basic search stays cheap but ten topics with full-page fetches on every source you find is a different scale entirely.
The Alternative Paths
Managed Agents isn't your only option. A few other paths work, depending on what's already in your stack:
Claude Code Routines does scheduling natively on Pro or Max.
You set the task and schedule and Anthropic runs it, but you're confined to the Routines interface for everything, including how the email gets delivered.
A VPS running Claude Code in headless mode with system cron gives you control over the whole pipeline.
Pipe claude -p "run briefing workflow" --allowedTools bash,web_search to your email script.
You also handle server maintenance yourself.
GitHub Actions with a schedule trigger works if YAML is already part of your day.
Free for public repos, and has capped minutes for private.
It is good for repo-adjacent jobs, less good for general information gathering since every run is a cold checkout.
I picked Managed Agents because it has API-level control, python instead of YAML, credential vault and sandboxed execution built in, and a container lifecycle handled by Anthropic.
My take
I've been running a version of this for a while, and it arrives at 7:04 AM every morning (the four-minute lag is session startup).
Two of my topics stay fixed (MCP ecosystem changes, Claude API updates), and the third rotates based on what has my attention that week.
My costs last month: about $2.10 in API spend, $0 in hosting, and the fifty lines of Python I wrote once haven't needed changes since.
Compared to the AI newsletter subscriptions I used to have, which ran $10–30 each and mostly covered things I half-cared about.
This one is exactly what I asked for.
Lastly, one thumb of rule: spend more time on the spec than the SDK if you're planning to point Managed Agents at your own use case.
The framework will run whatever spec you write but it can't tell you if the spec is the right one to solve your problem.
Until next time,
Vaibhav 🤝🏻
If you read till here, you might find this interesting
#Partner 1
Your first HR system, implemented right
Rolling out your first HR tool? Get a step-by-step guide to avoid common mistakes, drive adoption, and build a scalable HR foundation.
#Partner 2
Defense Tech Is the New Frontier for Investors
Record defense budgets are flowing into AI, satellite infrastructure, and advanced aerospace platforms. Our free research report names 5 companies at the center of this spending shift, before Wall Street catches up.
