A Proactive News Agent: HYPE, FACT, or SIGNAL

0:00IndyDevDan's thesis: in the age of AI there's "exponentially more noise than signal," and an engineer's time and attention are the most important resources. The meta lesson of this video is bigger than the tools: the best engineers are moving from human-in-the-loop → human-near-the-loop → human-out-of-the-loop. The demo is a proactive "durable news agent" that curates AI news to his taste, hands-off.

In → near → out of the loop

1:27Your agents can already do "80% or more of everything you're doing right now" autonomously if you have the right tech — but jumping straight from in-loop to out-loop is a big challenge. Proactive agents are the middle step: they don't wait for you to work — they work for you, then wait. When you're ready, you step in, extract value, give feedback, and move on.

What a proactive agent is

1:34The demo is a client-server app: a generative UI on the left, the agent on the right, both just clients — the real durable agent runs on his M4 Mac mini (and separately on an exc.dev sandbox). Browser, terminal, and CLI are all synced because they talk to the same agent through an API. Hands off the keyboard, a cron interval fires every minute and work happens whether he's there or not. The one-line definition: a proactive agent = a durable agent + a trigger, inside a sandbox.

Hype, Fact, Signal

5:10Before the architecture, he lays out the taste he's encoding — how he mentally separates new information into three buckets:

• Hype — too much attention, topic growth driven by unfounded excitement because the topic is attention-grabbing by nature.
• Fact — objectively true, but not necessarily actionable. Example: OpenAI dropping a huge set of math proofs — true and awesome, but not something an engineer can act on today.
• Signal — something you can actually act on. Example: Gemini 4 Argon rolling out to the API — a state-of-the-art model you and I can use right now.

That taxonomy is the applied judgment that shapes the system — "code is now infinitely cheap and infinitely copyable; what matters is your taste, your judgment, and your ability to architect unique systems."

The architecture

7:02Built in layers, from the bottom up:

• SQLite — durable storage (any database works).
• Pi Durable — the durable-agent implementation (shout-out to "Mario and Armin"): durable tasks triggered via an API, isolated, redeployable anywhere.
• A custom durable agent — his news agent built on top, exposed through an API.
• Jev — the TypeSafe AI classifier doing classification, filtering, and ranking at dirt-cheap prices.
• An SDK — what the browser, TUI, cron, and CLI all use.

8:30The point is an architecture that can be redeployed across many use cases "in a proactive agent-first way" — the same system runs on his M4 and on an exc.dev cloud sandbox, each with its own feed and settings. He pushes back on the "software engineering is solved" narrative: agents are a new tool, and "you can do more than ever now if you're using the right tools" — but you still need practitioners to drive the architecture.

The sweep workflow

10:07Each cron tick runs a sweep: gather → gate → dedupe → rank, returning N candidates; the durable agent then decides what to render from the rankings. Jev is used at nearly every step. 10:29Concretely, Jev answers: is this news piece relevant? is it a dupe / the same story? (fallback when a content hash can't be used — do the hash first), and is this claim actually fact? (synthesize many sources and check for overlap). It classifies large amounts of text in sub-half-second calls, dirt cheap and reliable — "roll your own if you want, I'm saving a ton of time using out-of-the-box Jev."

The data flywheel

12:29The key product feature is a data flywheel: a feedback loop between his inputs and how the code + agents improve. He upvotes/downvotes a card, and that's real signal fed back into the ranking algorithms. He can also talk directly to the agent — "increase the weight of Gemini; if it doesn't exist as a keyword, add it" — and it edits its own taste files (raising the Gemini keyword from 1.2 to 1.3), so all subsequent runs know what he cares about. The Jev classifier tags each story with its own hype/fact/signal mix (e.g. a model release scored "mostly signal, some hype — 66/100 hype because model labs always hype their releases").

The tools

15:46The agent's toolset is narrow and purpose-built (not a general code-editing agent): run news sweep, list candidates, read items, search web, scrape URL, research items, Jev classify, present news, present timeline, ask taste. All "harness-engineered tools" built with an in-loop agent to get the system running — "I'm building the system that builds a system." Cost stays flat because Jev and cheap models (Luna, Haiku 5.5, DeepSeek V4.1 Flash) gate the expensive calls. The closing reframe: 18:30durable agents are powerful because they make the in→near→out transition concrete — and that transition "is a direct measure of your agentic engineering skill."

Claims checked

Sources & further reading