I built an agent that researches AI news for me while I sleep.
Claude Managed Agents, a Memory Store, and one email in my inbox every few days. Here is why I built it and what I want it to become.
It feels like every day there is a new AI breakthrough. You go to bed and wake up to another one. After a while, keeping up starts to feel like a part-time job. So I decided to automate some of it.
Every morning I was spending too long checking various sources: Anthropic, OpenAI, Google DeepMind, xAI, Perplexity, Andrej Karpathy, trying to figure out what shipped overnight. It was not research. It was triage. I wanted something to do it for me.
What I built
A Claude Managed Agent that runs on Anthropic's infrastructure. No server to maintain, no containers to manage. I define what the agent does and Anthropic handles the rest.
The agent sweeps all six sources and produces a structured digest. Three sections every run: what shipped, what to try, and what might be worth a Field Note. Each item has a source link. Facts only, no opinion.
A GitHub Actions cron triggers the whole thing automatically. The digest arrives in my inbox before I open my laptop.
Why memory matters
Without memory, the agent has no idea what it already told me. The same story shows up every run. GPT-5.5 Instant becomes ChatGPT's default model on Monday. I hear about it again on Wednesday. Then Friday.
With a Memory Store attached, the agent checks what it already covered before each run and writes a brief summary of what it surfaced at the end. The next session reads that file and skips anything already sent. No database, no deduplication logic. Just a file the agent manages itself.
What I am waiting on
Two features are still gated: Dreaming and Outcomes.
Dreaming is a scheduled process that would automatically review past sessions and curate the memory store between runs. Right now the agent writes one history file per run and reads through all of them at the start of each session. Over time that gets unwieldy. Dreaming would consolidate everything automatically, extract patterns, and keep the memory clean without me touching anything.
Outcomes would let me define what a good digest looks like as a rubric. All six sources covered, items from the last 72 hours only, Field Note candidates with a specific angle not just a topic. A separate Claude instance grades each digest against that rubric before the email sends. If something fails, the agent gets specific feedback and takes another pass. Anthropic says Outcomes improved task success by up to 10 points in internal testing. The digest is already good. With a quality gate it would be consistently good.
Both are on my list. For now the pipeline runs, and that is enough.