Most of the conversation about agents concerns what they can do inside a terminal. The harder problems are at the edges, where a claim has to be trusted, a document has to be read, a video has to be produced, and several agents and people have to agree on what happened.
Eleven projects from the last stretch of Hacker News sit on those edges, and the pattern in them is more interesting than any individual tool.
Verifying what an agent tells you
Bullshit Detector is a set of agent skills that fact-check the internet, doing claim-by-claim verification with sources and returning a score from 0 to 10 for any YouTube video, article, tweet or PDF. Claim-level output is the part that matters. A single verdict on a long article is nearly useless, because a document can be ninety percent accurate and still lead you astray on the one line you act on. Separating the claims lets you check the specific one.
A related concern sits one layer down, in the model itself. Spanda provides zero-cost epistemic uncertainty quantification and hallucination detection without resampling, so you can attach an uncertainty estimate to a generation without paying for multiple samples. That is the difference between a system that can flag likely-bad answers and one that can only sample a few. It requires token-level information from the model, which makes it a library-level tool rather than a drop-in.
Strot addresses reliability from the compile-time end. It does compile-time LLM code synthesis with deterministic, self-routing NLU runtimes, so the intent is compiled once and the decisions run locally from then on. The model call happens at build time rather than on every request, which turns a variable per-call cost into a fixed build cost and makes the behaviour repeatable. It is very new, but the shape is one worth watching.
Reading documents and screens
An agent that cannot read a page cannot act on it, which produces two kinds of tool in this batch.
OCR It handles the case where text exists but cannot be copied. You pin a screen region once, then hotkey your way through a paginated document while OCR runs completely offline through a bundled Tesseract build. The pin-once workflow is designed for long documents, where reselecting a region on every page is the thing that makes the task miserable.
Accept Markdown attacks the same problem from the publisher's side. It documents serving a Markdown variant of a site to AI agents using Accept headers, so an agent client reads your prose instead of navigation, scripts and layout markup. The benefit splits two ways: fewer tokens spent on your page, and a cleaner embedding for retrieval pipelines because ads and related-content rails no longer pollute the text. For site operators who have noticed agent traffic and do not know what those agents are seeing, this is the direct answer.
Making video
Explainroo turns a script into an explainer video or product demo, free and open source. The pipeline is entirely local: narration from Kokoro, word timing from Whisper, and a canvas renderer producing a narrated MP4. Synchronising narration to what is on screen via word-level timings is the detail that makes it usable rather than a toy, and avoiding render credits entirely is the part that makes it deployable in CI.
The honest limitation is the medium. Canvas animation is not a screen recording, so it suits generated explainer content far better than capturing an existing interface in motion. For anything that needs live capture, this is the wrong tool.
Drawgent sits between reading and making. It is a coding agent that works on a live Excalidraw canvas, so a diagram and the code it describes can be edited by the same session instead of living in separate tools. Sketching architecture before coding is a common habit that agentic workflows have handled badly, because the sketch ends up as a screenshot nobody maintains.
Coordination between agents and people
Praxos is messaging built for the situation where you and your teammates each have several agent conversations running at once. It remembers what people and agents said and did, so a thread can be handed to Claude or Codex without copying messages over, and a colleague can pick up where an agent stopped. The failure it targets is real and unglamorous: the cost of re-explaining context from one conversation to the next.
Almanac approaches the same coordination problem as a workspace. It is a Mac app that keeps tasks, projects and knowledge in one place, combining a wiki, project boards and search, with the framing that agent-assisted work scatters across tools unless something holds it together.
Masala does it from the research angle. It is a research workspace where an agent works beside you, with chat, PDFs, notebooks and files in one dock, and custom panes the agent builds on request. Agent-built panes are the genuine novelty: instead of rendering whatever the model says into a text bubble, the agent constructs an interface for the thing it wants to show you.
Quant agents
TradingAgents wraps a multi-agent LLM financial trading framework, modelling a firm as specialised agents that analyse fundamentals, sentiment, news and technicals and then debate before deciding. It is widely used as research scaffolding rather than a trading system, and it is interesting here as a working demonstration of agent debate as an architecture: having several roles argue before a decision is made is a pattern that transfers well outside finance.
What these have in common
The unifying observation is that these eleven tools are all about the boundary of an agent's competence rather than its centre. Claim verification, offline OCR, Markdown negotiation, video rendering and multi-agent messaging are all places where a model's output would otherwise be wrong, unreadable, or unreachable by the next participant.
There is a general lesson. As models improve, the marginal value of anything that verifies, renders or relays the output goes up rather than down, because the outputs get more numerous and more trusted. None of these projects competes with a frontier model, and all of them would be useless without one.
