Pathfinder - Proposed Curiosity Agent Stack
Purpose
Before I spin anything up, this note explains:
- which agents I would create
- what each one does
- what you would actually experience as the user
- how the whole thing should feel in day-to-day use
The goal is not “more agents.” The goal is to create a small, coherent system that helps you:
- mine your existing notes for unexplored threads
- connect them with external technical trends
- maintain a living R&D frontier
- get high-quality prompts on demand when you feel bored or want direction
Design Principles
I would optimize the system for:
- Depth over hype
- Continuity over one-off novelty
- Synthesis over link-dumping
- High-conviction recommendations over large lists
- Usefulness in your real workflow rather than generic AI-demo behavior
That means the system should behave more like an R&D companion + second-brain maintainer than a generic trend bot.
Agent 1: Notes Distiller
What it does
This agent reads your notes and continuously builds a structured understanding of:
- recurring themes
- half-finished ideas
- unresolved questions
- abandoned but promising threads
- surprising connections across separate notes
It is the internal-facing memory and synthesis layer.
Inputs
- your Obsidian notes
- selected recent research notes
- prior topic-of-the-day outputs
- backlog of previously surfaced ideas
Outputs
It would maintain artifacts like:
interest-map.mdopen-loops.mdcuriosity-backlog.mdrecently-explored.md
Why it exists
Without this layer, the rest of the system becomes stateless and repetitive. It would keep rediscovering the same obvious interests instead of noticing the edges of your curiosity.
What you would experience
Mostly invisible background work.
Occasionally you would notice that I say things like:
- “You’ve circled around durable workflow execution three times across separate notes.”
- “There’s a half-explored thread in your notes around agent runtime isolation that looks worth reviving.”
- “You touched this topic before, but never pushed into the hard systems question underneath it.”
This agent makes the rest feel personal instead of generic.
Agent 2: Frontier Scout
What it does
This agent looks outward. It scans for external topics that may intersect your interests, such as:
- agent runtimes and orchestration patterns
- distributed systems research
- observability/tracing developments
- data infra and storage architecture shifts
- protocol/tooling changes
- interesting OSS projects
- meaningful technical debates
Inputs
- curated web sources
- papers / blog posts / repos / technical discussions
- trend signals in areas adjacent to your interests
- the internal interest map from the Notes Distiller
Outputs
For each candidate topic, it would capture:
- what the topic is
- why it matters technically
- why it is relevant to you specifically
- whether it is a new primitive, new pattern, new constraint, new application, or new disagreement
- suggested next angle to explore
Why it exists
A note-only system becomes introspective and stale. An internet-only system becomes shallow and hype-driven. This agent is the expansion layer that connects your existing thinking to what is moving in the world.
What you would experience
You would not get a giant feed. Instead, you would get occasional high-signal nudges like:
- “There’s a new runtime pattern here that intersects your agent orchestration interest.”
- “This trend looks noisy on the surface, but the underlying systems problem is real.”
- “This project is not broadly hyped, but it sits exactly one hop away from topics you’ve been exploring.”
Agent 3: Curiosity Ranker
What it does
This agent scores and prioritizes candidate topics from the first two agents.
Its job is to decide:
- what is genuinely interesting for you
- what is fresh enough to surface now
- what is deep enough to reward attention
- what is too similar to things you already explored
Scoring dimensions
I would likely score topics on:
- relevance to your real interests
- novelty relative to your recent exploration
- depth potential
- adjacency / cross-link value
- experimentability
- current timeliness
Why it exists
This is the anti-slop layer. Without it, the system becomes a topic spammer.
What you would experience
You’d see fewer but sharper recommendations.
Instead of:
- “Here are 12 interesting AI topics”
You’d get:
- “Here are the top 3. One practical, one conceptual, one speculative.”
This is the part that turns raw discovery into taste.
Agent 4: Topic Composer
What it does
This is the delivery agent. It turns the ranked backlog into things you can actually consume.
It would generate outputs in a few forms:
A. Topic of the Day
One strong recommendation with:
- title
- why this is interesting for you now
- 3 angles to explore
- optional hands-on experiment
- optional link to prior relevant notes
B. Boredom Queue
When you ask something like:
- “what should I focus on”
- “give me something interesting”
- “what rabbit hole do you have for me”
it returns up to 3 options:
- one practical
- one conceptual
- one speculative
C. Frontier Scan
A longer periodic summary that says:
- what unexplored topics were found in your notes
- what external developments look relevant
- what should be added to the backlog
- what should be retired as stale
What you would experience
This is the user-facing part. Most of your interaction would be with this layer.
It should feel concise, high-signal, and opinionated.
Optional Agent 5: R&D Program Manager
What it does
This is optional, but I think it could become very valuable.
Instead of just recommending topics, it would track:
- current research threads
- active experiments
- blockers
- next concrete step
- whether a thread should be delegated to another agent
- whether a thread should be paused or killed
Why it exists
There is a difference between:
- interesting things to think about
- and an actual personal R&D pipeline
This agent turns curiosity into momentum.
What you would experience
This would feel like a research chief-of-staff.
Examples:
- “You have 5 active threads. Two are alive, one is stalled, two should probably be archived.”
- “If you want a 30-minute task, do this. If you want a weekend rabbit hole, do that.”
- “This topic is interesting, but it doesn’t beat your current open loops.”
I would start without this agent, but it’s the most interesting second-phase upgrade.
Recommended Initial Stack
If I were starting conservatively, I would spin up 4 agents:
- Notes Distiller — internal note mining and topic extraction
- Frontier Scout — external discovery and trend intersection
- Curiosity Ranker — scoring, de-duplication, prioritization
- Topic Composer — user-facing outputs and notifications
That is enough to produce something useful without making the system bloated.
What You Would Experience as the User
From your side, the system should feel very simple.
1. Morning mode
Some mornings, if there is a strong candidate, you receive:
- a short “topic of the day”
- why it fits your interests
- 3 ways to approach it
- maybe a tiny experiment
Not every day by force. Only when there is something worth sending.
2. Bored mode
When you ask:
- “what should I look at?”
- “give me a rabbit hole”
- “what am I not seeing?”
I give you 1-3 tailored suggestions from the maintained backlog.
3. Review mode
Sometimes I give you a more reflective summary:
- unexplored themes found in your notes
- topics you keep orbiting
- external developments now intersecting your prior interests
- threads worth reviving
4. Quiet background mode
Most of the real work happens without bothering you:
- note mining
- trend scanning
- ranking
- de-duplication
- backlog maintenance
So the experience should feel lightweight even though the system behind it is not.
What It Should Not Feel Like
If built well, it should not feel like:
- a generic AI newsletter
- a hype-chasing trend bot
- a spammy notification engine
- a bookmark collector
- a random topic generator
If it ever starts sounding like “Top 10 trending AI ideas this week,” then we built the wrong thing.
Delivery Behavior I Recommend
Topic of the day frequency
I do not think this should fire every single morning no matter what.
Better behavior:
- send only when quality clears a threshold
- skip when the backlog has nothing strong
- maybe 3-5 times per week is better than forced daily sludge
Output length
The best default is probably:
- short in chat
- deeper note on demand
So the notification gives you the hook, and if you want depth, I expand it into a note.
Recommendation diversity
When giving multiple topics, I would intentionally diversify by type:
- practical systems angle
- conceptual/theory angle
- speculative/future angle
That keeps the system from becoming one-dimensional.
The Core User Commands I Imagine
From your side, you should be able to say things like:
- “What should I focus on?”
- “Give me a rabbit hole”
- “What unexplored topics do you see in my notes?”
- “What’s trending that actually intersects my interests?”
- “Give me something practical”
- “Give me something weird”
- “What old thread should I revive?”
- “What should be topic of the day tomorrow?”
And the system should already have the maintained state needed to answer well.
My Recommendation
I think the right path is:
Phase 1
Build the core curiosity system with these roles:
- Notes Distiller
- Frontier Scout
- Curiosity Ranker
- Topic Composer
Phase 2
If the system proves useful, add:
- R&D Program Manager
That gives us a clean progression:
- first make it good at surfacing ideas
- then make it good at managing your personal research pipeline
Final Take
If I spin this up, I would not frame it as “some AI agents that recommend stuff.”
I would frame it as:
a long-running curiosity and R&D system that maintains a live model of what you care about, what you’ve already explored, what remains underexplored, and what the external frontier is doing nearby.
That is the version worth building.
Next Step
Before implementation, the next thing to lock down is the exact system prompt / contract for these agents:
- their responsibilities
- shared memory/state
- output formats
- notification rules
- ranking policy
- anti-slop constraints
That is the part I would design next.