How I Run A Full-Blown AI Research Operation on My Phone (Powered by Claude Cowork)

Dispatch for mobile access, Scheduled Tasks for daily briefings, and a knowledge system that gets smarter every day.

Mar 19, 2026

∙ Paid

AI research intelligence agent built with Claude Cowork — a compound knowledge system that scans the internet, tracks trends, and delivers daily AI briefings that get smarter over time

A few months ago, I built an AI agent that sends me AI news summaries every week. Perplexity searches the internet. Make.com orchestrates the pipeline. OpenAI writes the summary. Gmail delivers it. Set it and forget it.

It works. It still runs. It saves me 3-5 hours a week.

And I’ve been slowly realizing it’s not enough.

Make.com AI news automation pipeline showing the linear workflow — Perplexity searches, OpenAI summarizes, Gmail delivers — that runs weekly but starts from zero every time

Week after week, the same format lands in my inbox. Solid summaries. Useful links. Zero awareness of what it told me last week. The agent doesn’t know that three companies announced agent frameworks in the past month because it doesn’t remember the first two. It doesn’t know I already wrote about Claude Code because it has no logs of my newsletter archive. It can’t tell me “this contradicts what Source X reported last Tuesday” because last Tuesday doesn’t exist to it.

Every run starts from zero. Fresh amnesia. Great at the task, zero accumulated knowledge.

The email arrives. That’s it. If something catches my eye and I want to go deeper? I’m back to manual research. I can’t ask follow-up questions to a Make.com scenario.

What I actually wanted was a research partner, something that remembers every finding, understands what matters to me, and gets better at the job over time. Not a pipeline. An agent.

Same goal. Completely different architecture.

Side-by-side comparison table of Make.com AI agent versus Claude research agent showing differences in architecture, memory, context, pattern detection, and cost

If you built V1 from my previous post, this is the upgrade. If you didn’t, this system stands on its own.

Inside the AI Research Agent: Architecture, Modes, and How It Works

Cowork AI research agent system overview showing five capabilities — Dispatch for mobile remote control, Scheduled Tasks for automatic daily briefings, Subagents for parallel research, Tavily MCP for live internet access, and three-layer knowledge system for compound intelligence

An AI automation follows the same steps every time regardless of what happened before. An agent remembers what it learned and adapts.

We’re going to leverage two of Claude Cowork’s most recent features to make this work: Dispatch and Scheduled Tasks.

If this is your first time hearing about Claude Cowork, you may want to read my full guide on how to use Cowork.

The Ultimate Guide to Building Your Agentic AI Workflow With Claude Cowork

Wyndo

Feb 19

Read full story

Under the hood, the system combines five capabilities:

1. Dispatch

This is Cowork’s mobile entry point is worth pausing on. Dispatch gives you the full power of Cowork from your phone. Everything Cowork can do on your desktop—spin up subagents, read and write to local folders, connect to MCP servers like Gmail, Google Drive, Slack, etc., execute files—Dispatch can trigger from a text message on your phone.

Imagine you’re on the train, waiting for coffee, sitting in a boring meeting—pull out your phone, tell Claude “research autonomous coding agents,” and walk away. Claude picks it up on your desktop, deploys parallel subagents, scans the internet through Tavily MCP, writes the finished report to your local folder, and it’s all waiting when you get back. Your phone is the remote control. Your desktop is the engine.

And you know what? The more I use Dispatch, the more I realize this is Anthropic’s answer to OpenClaw, letting you access a full-blown AI agent anywhere. Even though Anthropic launched remote control in Claude Code a while back, having it in Cowork unlocks more possibilities and gives many more people access to it.

A tweet by Felix Rieseberg Anthropic showing how Claude Dispatch can trigger Claude Code sessions

You know what makes this exciting is that less than 24 hours ago, Anthropic made this 10x better by letting Dispatch launch a Claude Code session directly inside your computer.

Now you can ask Dispatch to open Claude Code in your terminal and start building for you. You can also connect to Google Workspace CLI through Dispatch via Claude Code. In a future post, I’ll do a deeper dive on this, so stay tuned.

2. Scheduled Tasks

This is the recurring layer. Schedule your daily briefing to run at a set time while your computer is on. You step away to grab lunch, come back, and the intelligence report is sitting in your folder. You don’t need to prompt or trigger anything. The agent just runs on schedule and the results accumulate.

3. Subagents

Claude can deploy multiple parallel agents that each handle a different research topic simultaneously. Instead of searching one thing at a time, the system fans out across all your tracked topics at once and then synthesizes the results.

4. Tavily MCP

Tavily product capabilities: search, extract, research, crawl, and map

This is the agent’s connection to the live internet. Tavily offers five capabilities: search to scan for new developments, extract to read full articles without noise, research to run comprehensive multi-source investigations, crawl to capture entire sites, and map to understand site structure before crawling. The free tier covers daily briefings. Pay-as-you-go if you run heavy deep research sessions.

5. Three-layer knowledge system

This system separates what the agent already knows about you, what it is actively tracking, and the day‑to‑day details that eventually compound into something useful.

This system has three modes:

Three-layer knowledge system diagram showing how the AI research agent separates memory (user preferences), trend index (accumulated intelligence), and daily logs (session detail) to compound research over time

Daily Briefing (5-10 minutes) — You say “morning briefing.” Claude reads its memory (your preferences), checks the trend index (everything it’s been tracking), reads yesterday’s log (what’s pending), reviews your research profile (what you care about) — then deploys parallel subagents to scan all your topic areas simultaneously. It synthesizes findings, connects them to previous days, and delivers a focused report. The briefing on Day 30 looks nothing like Day 1 — because the agent has 29 days of accumulated intelligence informing what matters and what’s noise.
Deep Research (10-15 minutes) — You say “research [topic].” Claude goes deep. But unlike a fresh search, it starts from everything it already knows. It checks the trend index for existing data points, then runs a comprehensive multi-source sweep before deploying parallel subagents to fill specific gaps. If you’ve been tracking a trend for two weeks through daily briefings, the deep research builds on that foundation instead of starting cold. This one takes longer because it’s thorough: multiple research phases, gap-filling, confidence scoring.
Trend Review (5-10 minutes) — You say “trend review.” Claude pulse-checks every active trend it’s tracking, makes explicit decisions (keep, promote, archive, flag for deep dive), identifies meta-patterns across trends, and produces a landscape summary. This is the system’s strategic assessment mode — it keeps the accumulated intelligence healthy and surfaces the big-picture view that daily briefings miss. Run it weekly.

Here’s what that looks like in practice:

Day 1: “Here are today’s top AI developments.” (Clean scan. No context yet.)
Day 5: “This connects to Tuesday’s finding — OpenAI is the third company this week to announce agent frameworks.” (Pattern recognition kicks in.)
Day 15: “Based on two weeks of tracking, here’s what’s actually signal vs. noise.” (Accumulated intelligence.)
Day 30: “You asked me to research AI coding tools. Based on everything we’ve tracked, here’s what’s shifting vs. what’s just marketing.” (Deep research powered by a month of context.)

The whole thing lives in one folder:

Open this folder in Cowork and type “morning briefing.” That’s it. (If you prefer Claude Code, run claude in the directory — same files, same system.)

How the Agent Gets Smarter: Day 1 vs. Day 7

Let me walk you through what actually happens when you use this system for a week. These are real outputs from my own research agent, not mockups.

Day 1: A Good Summary

Claude Cowork interface showing the research agent folder attached and morning briefing command typed — one prompt to trigger the full AI research agent using Opus 4.6 with 1M context

I open Cowork, point it to my research-agent folder, and type “morning briefing.”

The agent kicks off. It reads memory, trend index, yesterday’s log (all empty — first day), and my research profile. Then it deploys parallel subagents — one per topic area in my profile — each scanning the internet using Tavily MCP simultaneously.

Five parallel subagents running simultaneously in Claude Cowork — each scanning a different research topic area at once, then cross-referencing and synthesizing findings into a single daily briefing

A few minutes later, the first briefing lands.

AI research agent daily briefing output showing Top Signal on MCP production roadmap and five key developments including Anthropic multi-agent code review, OpenAI GPT-5.4, Anthropic labor market report, Claude inline visualizations, and Shadow AI adoption data

Here’s the structure of what Day 1 delivered:

Top Signal: MCP’s 2026 production roadmap — security deliberately deprioritized despite a major vulnerability that compromised 437K+ dev environments. The agent flagged the tension: the protocol everyone’s building on doesn’t yet enforce security at the protocol level. Strong newsletter angle.
5 Key Developments: Claude Code Review (multi-agent architecture), GPT-5.4 with computer use, Anthropic’s labor market report, Claude inline visualizations, shadow AI hitting 57% employee adoption
Emerging Patterns: Empty — first briefing. “No patterns identified yet. This section populates as context accumulates — typically by Day 3-5.”
Worth Watching: 4 items on the radar for future tracking
Memory Updates: “No changes — first briefing, no user preferences to record yet.”

Behind the scenes, the agent also:

Claude Cowork task screen showing file access

Seeded the trend index with 6 emerging trends from today’s findings
Created a detailed daily log with every source consulted and follow-up questions
Saved the briefing to output/daily-briefings/2026-03-17.md

The overall result was solid and clean. This is a good summary of today’s landscape.

Day 7: When the Agent Started Connecting Dots

Same command. “Morning briefing.” But now the agent has a week of logs and a trend index carrying six active trends with dated evidence trails.

Emerging Patterns section from AI research agent briefing showing compound intelligence in action — dated evidence trails across multiple days tracking cognitive offloading research and workspace AI integration with confidence levels and cross-source assessments

The briefing is different in ways I didn’t design for. Two things happened that week that made me realize this wasn’t just a faster news reader:

It connected dots I missed. Three separate announcements across the week, each one looking like independent news, turned out to follow the same pattern: they were all about workspace AI integration. The agent saw it because it remembered all three. I would have read them in three separate newsletters and never connected them.
It caught a contradiction. The agent flagged that something a company announced directly conflicted with what a different source reported earlier in the week. The trend index carried both data points with dates. It surfaced the tension unprompted: “This conflicts with [Source]’s earlier report.” That contradiction became the angle for a newsletter post I wouldn’t have written otherwise.

The system is measurably smarter than it was on Day 1. And next week, it starts from here, not from zero.

Everything Inside the Research Agent Folder

That’s what the system does. Now here’s what you’ll walk away with if you keep reading.

The complete research agent folder: every file, configured and ready to use. Drop it into a directory, open Claude, and type “morning briefing.” That’s it.

Here’s what’s inside:

CLAUDE.md — The brain. 250 lines of operating instructions that turn Claude from a chatbot into a research intelligence agent. Includes the three-layer knowledge system, a full Tavily MCP usage guide (when to use search vs. extract vs. research vs. crawl vs. map), parallel subagent architecture for concurrent topic scanning, proactive suggestion logic, and the exact session startup sequence the agent follows every time.
research-profile.md — Where you tell the agent who you are and what you care about. Your topics, your trusted sources, your filters for what matters and what’s noise. I’m including 3 complete example profiles (creator, knowledge worker, entrepreneur) so you can start from a template instead of a blank page.

3 skill files — The step-by-step processes for each mode:

Daily Briefing (10 steps, parallel subagent deployment, proactive suggestions)
Deep Research (11 steps, two-phase research with gap-filling)
Trend Review (10 steps, keep/promote/cold/archive decisions)

3 output templates — Consistent formatting for briefings, research reports, and trend reviews so every output is scannable and structured.

The three-layer knowledge system — Starter files for memory.md (user preferences), logs/trend-index.md (accumulated intelligence), and the daily log format. Each with commented examples showing exactly what goes where — and why the separation matters.

Plus the full setup and tuning guide:

Step-by-step setup walkthrough (Cowork primary, Claude Code alternative)
Tavily MCP connection setup (2 minutes by giving the agent live internet access)
How to build your research profile inside Cowork — let Claude interview you and write the file for you
Dispatch setup: trigger research from your phone, results waiting on your desktop
Scheduled Tasks: automatic daily briefings that run without you prompting
What every file does and why it exists (so you can tune the system with confidence)
Compound intelligence timeline (what to expect at Week 1, 2, and 4)
Week 1 tuning checklist — what to adjust after your first 5-7 briefings
How your source list grows naturally over time
Honest results: what improved, what didn’t, and what’s still imperfect

What you need: Claude Pro ($20/month) or Claude Max — Cowork is the easiest way to run this (Claude Code works too). Tavily MCP for internet research (free tier of 1,000 credits/month is available, with pay‑as‑you‑go pricing at $0.008 per credit for heavier usage). That’s it. You don’t need a Make.com subscription, an OpenAI API key, or even a multi-tool pipeline to maintain.

The whole thing takes about 15 minutes to set up. The research profile is the only part that requires real thought — everything else is copy, paste, and go.

Day 1 will feel like a good news summary. Day 5 will feel like having a research partner. Day 30 will feel like having an intelligence operation.

How to Set Up Your AI Research Agent (Step by Step)

Step 1: Download the Agent Folder and Connect Tavily MCP

Download the complete research-agent folder (link below). Drop it anywhere on your machine. Every file is ready to use, you only need to customize one.