Build a Research Agent That Sends You a Weekly Topic Digest

The Problem With Staying Current on a Topic

Most people track a subject the same way: a pile of RSS feeds they skim unevenly, a subreddit they sometimes visit, a newsletter that mixes signal with noise, and a vague intention to search "what's new in [topic]" that rarely happens. The result is either information overload or a weeks-long blind spot.

A research agent solves this cleanly. You configure it once — topic, sources, output format — and it does the weekly reading for you, writes a summary of what matters, and delivers it where you already check in. No dashboard to log into, no manual curation session to schedule.

This guide walks you through exactly how to build one: the logic, the prompt design, the memory strategy, and how to have it running in minutes on AgentRoost with no server work.

---

What the Agent Actually Does

Before the build steps, here is the full loop in plain terms:

1. Scheduled trigger — runs once a week (e.g., every Monday at 07:00).
2. Source fetch — queries a set of URLs, RSS feeds, or a search API for recent content on the topic.
3. Deduplication check — compares incoming items against a persistent memory store of already-sent links.
4. Summarisation — feeds the new items to an LLM with a prompt that extracts the key insight from each in 2–3 sentences.
5. Digest assembly — formats the summaries as a numbered list with titles and links.
6. Delivery — sends the finished digest as a Telegram message.
7. Memory update — writes the newly sent URLs back to persistent memory so they are never surfaced again.

Steps 3 and 7 are what separate this from a simple scheduled search. Without them you get the same articles every week.

---

Designing the Prompt

The quality of the digest lives in the summarisation prompt. A weak prompt produces vague, interchangeable summaries. A strong one makes each bullet actually useful.

A prompt that works well:

You are a research analyst tracking developments in: {{TOPIC}}.

You have been given a list of articles published in the last 7 days.
For each article:
- Write a 2–3 sentence summary focused on what is NEW or CHANGED.
- Note why it matters for someone watching this space.
- Include the original URL.

Skip any article that is an opinion piece with no new facts.
Skip any article you have seen before (the seen list is in your memory).
Format as a numbered Markdown list.

The {{TOPIC}} placeholder gets filled at runtime. The "seen list" reference is what triggers the agent to check its memory before writing.

You can tune this prompt in two ways as you go:

• Tighten the scope: "Focus only on developments related to regulatory changes, not product announcements."
• Adjust the voice: "Write for a non-technical founder, not an engineer."

---

The Memory Layer: How Deduplication Works

Persistent memory is the feature that turns a one-shot summariser into a tireless analyst. The agent maintains a simple running log — essentially a list of URLs it has already included in a past digest.

Each time it runs:

1. Fetch candidate articles.
2. Load seen_urls from memory.
3. Filter: remove any article whose URL is in seen_urls.
4. Summarise the remainder.
5. Append the newly used URLs to seen_urls and save.

This is intentionally simple, and that is the point. The LLM does not need a vector database for this use case — a structured list in the agent's persistent state is enough. The Hermes framework on AgentRoost maintains this state across restarts, reschedulings, and updates without any infrastructure work on your part.

For a topic where you care about concepts rather than exact URLs (e.g., you want to avoid reporting the same event from multiple angles), you can extend the memory to include a short string description of each story: "OpenAI announced o3 pricing — June 2025." The agent checks that list before summarising.

---

Source Strategy

The agent needs something to read. The right sources depend on the topic:

Topic type	Good sources
Tech / AI	RSS from Hacker News, arXiv daily digest, specific substack feeds
Market research	Google News RSS for a search query, industry publication RSS
Competitor tracking	Their blog RSS + mentions via search API
Regulatory / legal	Government RSS feeds, law firm newsletters
Niche communities	Subreddit RSS (reddit.com/r/[sub]/.rss), Lobsters RSS

Reddit and most major news sites expose RSS at predictable paths. A few sources (Twitter/X, LinkedIn) do not — for those you need a search-based approach rather than a direct feed. An HTTP GET to a news search endpoint with a date filter (after:7d) is a reliable fallback.

Start with two or three sources. More sources means longer runs and bigger LLM context. Add more only once the digest quality is solid.

---

How to Run This on AgentRoost

The Hermes framework is purpose-built for exactly this pattern: persistent state, scheduled tasks, Telegram delivery, and included AI credits. Plans start at $19.99/month all-in — that covers the compute, the Telegram bot, and the AI credits the summarisation step consumes. Here is the path from zero to a running agent.

Step 1 — Sign up

Go to agentroost.app and create an account (email/password, or Google, Microsoft, or Discord).

Step 2 — Create a new agent

From your dashboard, choose "New Agent" and select Hermes as the framework. Give the agent a name — something like weekly-research-ai-news.

Step 3 — Configure topic, sources, and schedule

In the agent configuration, you set:

• The topic string that fills {{TOPIC}}.
• The list of RSS/HTTP source URLs.
• The scheduled trigger: every Monday at 07:00 UTC (or whatever cadence you want).
• The delivery channel: Telegram (Hermes provisions a bot automatically).

Step 4 — Connect Telegram

Open the AgentRoost manager bot on Telegram and send /start to your new agent. The bot comes alive. No BotFather, no tokens to copy, no webhook URL to configure — it is handled.

Step 5 — AI credits are already included

This is where AgentRoost differs from every other hosting option. On Elestio, Sliplane, or any self-hosted setup, you would be copying an OpenAI key into the LLM node right now. On AgentRoost, the AI credits are already included in your subscription. The summarisation step uses whichever of the 350+ available models you specify — the default works well for research digests, or you can pick a model optimised for long-context document reading.

Step 6 — Test it

Trigger a manual run before the first scheduled run. The digest arrives in your Telegram within a couple of minutes. Check the summaries for quality, adjust the prompt if needed, and let the schedule take over.

The whole setup — from sign-up to first digest — takes about 2 minutes once you have your sources and topic ready.

Compare plans and get started →

---

Tips and Pitfalls

Pitfall: Sources that block bots. Some sites return 403s to automated fetches. Use their RSS feed instead of scraping the HTML. If no RSS exists, route through a search API with the site as a filter.

Pitfall: Context window overflow. If you feed 30 articles into a single LLM call, the context gets noisy. Batch in groups of 8–10 articles, summarise each batch, then combine the batch summaries into the final digest. This also makes the output length predictable.

Tip: Keep the memory lean. You do not need to store full article text — just the URL and optionally the headline. The memory list stays small and lookups stay fast.

Tip: Add a "highlights" section. Ask the agent to prepend a 2-sentence "This week's most important development" before the numbered list. It forces a quality ranking and makes the digest faster to skim.

Tip: Send a "nothing new" message. If all fetched items are already in memory, have the agent send a brief "No new developments this week" rather than silence. Silence is ambiguous — did it run or not?

---

What You End Up With

A tireless analyst that reads your sources every week, remembers what it already told you, and delivers a clean numbered summary to your Telegram — powered by included AI credits, no server to maintain, no API keys to rotate, no cron job to babysit.

If you also want a research agent you can have a back-and-forth conversation with — ask follow-up questions, request deeper dives — the OpenClaw framework pairs naturally with the Hermes digest pattern.

For a broader look at what you can automate, explore all agent frameworks →.