The Scheduled AI Chief of Staff
Put proactive agents on a schedule to watch your metrics, surface misalignment, and coach you weekly
- Difficulty
- Moderate
- Time to result
- ~days to results
- Steps
- 5
- Confidence
- 80%
Rather than only prompting AI reactively, Amol runs standing scheduled agents that proactively do the work of a chief of staff. A co-work task reviews 20-25 charts every morning and flags what's concerning; another scans Slack weekly for cross-functional misalignment; another builds a model of a colleague or manager from their writing and Slack activity and returns the feedback they'd likely give. Trust builds as the false-positive and false-negative rates fall — buying peace of mind on the long tail of things you can't check daily.
Origin
Amol Avasare's personal working system at Anthropic, built on co-work (desktop app), scheduled tasks, the Chrome extension, and the Slack MCP; a concrete instance of the emerging 'strategy bot' pattern.
Core principles
- 01Shift AI from reactive prompting to proactive, scheduled monitoring of the long tail you can't watch daily
- 02Point agents at your real context — dashboards (via Chrome extension/MCP), Slack, transcripts, OKRs — not generic prompts
- 03Trust is earned empirically: watch the false-positive and false-negative rates fall before relying on it
- 04Offload life-admin (meeting rooms, inbox triage, expense filing) entirely to reclaim attention
- 05A public writing corpus or Slack history lets an agent model a person well enough to predict their feedback
How to run it
- 1
Set up co-work with the right connections
Download the co-work desktop app and connect the data sources — the Slack MCP (may require team/enterprise admin permission), the Chrome extension, and links to your dashboards (e.g. hex links).
Pro tip Once the Slack MCP is connected in co-work, you often 'just ask Claude' — no elaborate setup required.
Watch out Connecting org-wide Slack may need an admin with the right permissions to enable it.
- 2
Schedule a morning metrics review
Create a scheduled co-work task that reviews your 20-25 charts every morning and returns what to pay attention to, what's concerning, and interesting insights — so you triage the important charts instead of all of them.
Pro tip Keep looking at a few charts yourself to sanity-check the agent and stay calibrated.
- 3
Schedule a weekly misalignment scan
Have the agent look across Slack — given the projects you're working on and what's top of mind — and surface areas of potential misalignment. Run it on a weekly schedule.
Pro tip This same pattern scales: point it at more data sources to approach a full 'strategy bot' watching metrics, market, and roadmap.
- 4
Build a coaching model of key people
For a manager or report with a public writing corpus (or rich Slack history), have the agent study it plus your discussions, team goals/OKRs, and transcripts, then ask what feedback they'd give you — or what feedback you should give them.
Pro tip Amol runs this weekly both for his reports and for himself, modeling his own manager to get proactive coaching.
Watch out Quality is hit-or-miss — treat it like a coach who is occasionally wrong; verify before acting on surprising claims.
- 5
Calibrate trust before offloading
Track whether the agent's false-positive and false-negative rates are dropping week over week. As they fall, extend how much you rely on it — including handing off pure life-admin like booking rooms, inbox first-pass, and filing expenses.
Pro tip You can always unship or ignore a bad suggestion, so the downside of trying is low.
In the wild
Scott, who leads Anthropic's enterprise team, used AI scanning across context to find major areas of misalignment that would otherwise have caused teams to spin their wheels or do overlapping work.
→ Amol frames this as AI reducing cross-functional coordination toil that constrains shipping at a big company — work he says wasn't possible six months earlier.
Common mistakes
Acting on agent output before trust is calibrated
Coaching and misalignment output is hit-or-miss and sometimes 'clearly wrong'; relying on it before watching false-positive/false-negative rates fall leads to acting on bad signal.
Keeping AI purely reactive
Only prompting on demand misses the long and medium tail of metrics and misalignment you don't have time to check daily — the value comes from proactive, scheduled monitoring.
Is it for you?
Best for
Operators and managers drowning in dashboards, Slack, and cross-functional context who use co-work/Claude with MCP access to their real tools
Not ideal for
People without access to the underlying data connections (Slack MCP, dashboards) or in orgs where admin won't enable them, and anyone who would act on agent output without verification
From the transcript
“I personally have like a coworker runs on a schedule and looks at sort of 20 25 different charts every morning”
“go and find me areas of potential misalignment right now. And it does a really, really good job. So, this is something that I have…”
“what feedback do you have for me as Army?”
From the episode
Head of Growth (Anthropic): “Claude is growing itself at this point”
Amol Avasare