Skip to content

Anthropic

Anthropic’s daily-brief agent treats a failed read as news

6 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: x.com

Anthropic’s October 8 guide to the Claude Managed Agents beta builds a daily-brief agent that ends its Slack post with a line like “pull requests unavailable this run” whenever a source fails, so an expired token shows up in the brief instead of passing for a quiet day. The agent reads Slack channels and GitHub pull requests on a cron schedule, posts to one Slack channel from Anthropic’s infrastructure and picks up each source where the previous run stopped.

Where each run picks up

Each bookmark is the timestamp of the newest item the agent read from that source, written at the end of every run to one file, bookmarks.json. It replaces what the guide calls a common mistake, a fixed window like “the last 24 hours”: a late run leaves a gap and an early one repeats items, while a bookmark stretches or shrinks the window to cover everything since the last run. The file lives in a memory store called state, a folder of text files the platform mounts into every run’s sandbox under /mnt/memory/ and keeps between runs, since each run otherwise starts fresh.

If an MCP server is down or its token has expired, the run still starts, just without that server’s tools. The platform only logs an error in the session, and the agent, seeing nothing from that source, reports “nothing new.” Three rules in agent.md handle the case: keep that source’s bookmark where it is, write the brief from the other sources, and end it with one line naming what couldn’t be read.

Because the bookmark stays put, the next run that reaches the source covers everything since its last good read. An expired GitHub token is replaced in the vault, and running sessions pick up a changed key on their next call, Anthropic said in its June 9 announcement of scheduled deployments.

What counts as posted

The agent touches its ledger of reported items and its bookmarks after a post is confirmed, never before, and agent.md defines confirmed in three steps. Before posting, the agent looks for today’s title among the channel’s recent messages and stays quiet if the edition is already there. The post counts as sent only when Slack returns "ok": true and a message ts. When the result is unclear, the run is marked “maybe posted” and nothing else changes.

A post recorded but never delivered moves the bookmarks on, so its items are never reported; a second attempt by an agent unsure whether the first landed gives readers the same brief twice. A run record in memory, runs/<date>.md, goes from “posting” to “posted” with the message ID, or to “maybe posted.”

Just before posting, the Verify step re-checks every item against its live source: an item resolved since it was read is dropped, one that changed gets its line fixed, and one that can’t be confirmed is dropped and listed in the run record’s cuts. Links are copied from each source’s own link field, such as a pull request’s html_url or a Slack permalink. The step’s own text gives the reason: “One stale ‘still waiting on you’ costs more trust than ten missing items.”

The Decide step before it sets the bar for a line: the reader would act on the item today, or it changes a decision they are about to make; when unsure, the instruction says, “leave it out.” A count like “12 open reviews” is not an item, and something still open from an earlier brief comes back as one marked line, “still waiting, day 3.”

What runs without approval

MCP tools ask for approval by default, and on a scheduled run nobody is there to give it, so the template sets the GitHub toolset to always_allow and pairs it with a read-only token. The built-in bash tool runs without asking by default, which leaves the Slack post, a single curl call to chat.postMessage, going out unapproved with the same bot token the agent reads with; slack.com is also on the allowlist of the agent’s environment. Web search and web fetch are switched off, and each run has a spending cap set in the deployment.

The real credential values stay in a vault, outside the sandbox where Claude’s code runs. GitHub calls go through an MCP proxy outside the sandbox, which picks the vault credential whose URL matches the server’s. Inside, the Slack token exists only as an opaque placeholder, $SLACK_BOT_TOKEN, and the platform swaps in the real value as the request leaves, for hosts you allow.

Your preferences, from the channels and repos to read to the length cap, sit in a second memory store that the deployment mounts read_only with the instruction “Never write here.” A read_only mount rejects writes at the filesystem level, Anthropic’s memory documentation says, so the platform enforces that instruction. Changes the agent wants to your preferences go into its own state store as proposals.

Setting it up

In Claude Code, the claude-api skill configures the agent through /claude-api managed-agents-onboard followed by the guide’s URL; you need a Slack app created from the repo’s manifest and a GitHub token. The repo holds configuration files for the agent, its environment, both memory stores, the vault and the deployment. The ant apply command creates them in your Claude API workspace and records their IDs in claude-lock.json; Anthropic added it to its ant CLI on September 4, and with it a team can keep a whole Managed Agents setup under version control. Credentials go into the vault through the TypeScript SDK.

The template runs claude-sonnet-5-5 on the cron expression 32 7 * * 1-5 in America/New_York, 7:32 a.m. on weekdays. The time zone goes in twice: the deployment’s timezone field sets when the run fires, and the second line of its first message tells the agent which zone to compute dates in. Without that line, the agent can call this morning “yesterday” by computing dates in the server’s time zone, a bug the guide calls common.

Managed Agents bills standard Claude Platform token rates plus $0.08 per session-hour of active runtime. Anthropic’s ClaudeDevs account on X adds that the new API credits included in Max and Team plans can be used to run it.

To test it without waiting for the schedule, start a run by hand with ant beta:deployments run --deployment-id <id>, taking the ID from claude-lock.json.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.