Feed aggregator

Article URL: https://github.com/iii-hq/agentos

Comments URL: https://news.ycombinator.com/item?id=47304963

Points: 1

# Comments: 0

Article URL: https://www.economist.com/finance-and-economics/2026/03/05/to-understand-why-countries-grow-look-at-their-firms

Comments URL: https://news.ycombinator.com/item?id=47304939

Points: 1

# Comments: 1

Article URL: https://videocardz.com/newz/der8auer-explains-thermal-grizzly-delays-after-company-was-scammed-twice-on-raw-materials-worth-e40000

Comments URL: https://news.ycombinator.com/item?id=47304934

Points: 2

# Comments: 0

Article URL: https://twitter.com/alexwg/status/2030217301929132323

Comments URL: https://news.ycombinator.com/item?id=47304800

Points: 1

# Comments: 0

Article URL: https://cfp.3mdeb.com/qubes-os-summit-2024/talk/FCENX9/

Comments URL: https://news.ycombinator.com/item?id=47304796

Points: 1

# Comments: 0

Article URL: https://github.com/webcoyote/sandvault

Comments URL: https://news.ycombinator.com/item?id=47304786

Points: 1

# Comments: 0

Article URL: https://www.robpanico.com/articles/display/?entry_short=the-wrapper

Comments URL: https://news.ycombinator.com/item?id=47304774

Points: 1

# Comments: 0

Article URL: https://github.com/AnonJon/kroot

Comments URL: https://news.ycombinator.com/item?id=47304755

Points: 1

# Comments: 1

An "awesome list" with a twist: every entry is a verified case study, not just a link. To appear in this catalog, an agent must pass tracecore run --strict-spec in public CI — producing an immutable, schema-validated artifact as evidence.

GitHub Actions is the public gate. No human approval alone can certify an agent; the workflow must pass first.

Comments URL: https://news.ycombinator.com/item?id=47304740

Points: 1

# Comments: 1

Article URL: https://euclid.tulv.in/

Comments URL: https://news.ycombinator.com/item?id=47304738

Points: 1

# Comments: 0

Article URL: https://circuitcellar.com/research-design-hub/projects/executive-decision-maker/

Comments URL: https://news.ycombinator.com/item?id=47304717

Points: 1

# Comments: 0

Article URL: https://github.com/felipemarinho97/magnet-metadata-api

Comments URL: https://news.ycombinator.com/item?id=47304666

Points: 1

# Comments: 1

MongoDB is great until you read the SSPL. Then you're either paying Atlas prices, running an old 4.x build, or pretending FerretDB is production-ready. We built a third option.

Salvobase is a MongoDB wire-protocol-compatible document database written in Go. Point any Mongo driver and it works. No driver changes, no config changes. It's Apache 2.0, so you can embed it in a commercial product without a legal conversation. What it does: - Full CRUD, indexes (single, compound, unique, text, TTL, partial, wildcard), and most of the aggregation pipeline ($match, $group, $lookup, $unwind, $facet, etc.) - SCRAM-SHA-256 auth - bbolt storage engine: one .db file per database, Snappy-compressed BSON - Built-in Prometheus metrics at :27080/metrics (no exporter needed) - Built-in REST/JSON API at :27080/api/v1/ (MongoDB's equivalent is paid Atlas) - Per-tenant rate limiting, audit logging, 1-second TTL precision, SIGHUP hot reload - make build && make dev and you're running What it doesn't do: No replication. No sharding. No change streams. No multi-document transactions (stubbed). No $where or mapReduce (intentional: security + complexity). Single-node only. If you need a distributed MongoDB replacement, this isn't it yet. But we hope one day it will become that, built by agents. The weird part: The codebase is maintained by AI agents. Not "AI-assisted" - the agents pick issues from the backlog, write code, submit PRs, review each other's PRs, and merge. There's a formal protocol (https://github.com/inder/salvobase/blob/master/AGENT_PROTOCOL.md) covering identity, trust tiers, anti-collusion rules for reviews, claim timeouts, and a kill switch. Humans set direction; agents do the execution. We're curious whether autonomous agent maintenance can sustain a real open source project over time, not just generate initial code. * If you want to donate an agent just drop this prompt into Claude Code, Cursor, Aider, Devin, whatever: Fork/clone github.com/inder/salvobase, read QUICKSTART.md, and start contributing

GitHub: https://github.com/inder/salvobase

Thank you.

Comments URL: https://news.ycombinator.com/item?id=47304607

Points: 1

# Comments: 0

Article URL: https://rocketgraph.app/ml

Comments URL: https://news.ycombinator.com/item?id=47304597

Points: 1

# Comments: 1

Article URL: https://datarepublican.substack.com/p/data-analysis-of-the-state-of-the

Comments URL: https://news.ycombinator.com/item?id=47304592

Points: 1

# Comments: 0

Article URL: https://www.psychologytoday.com/us/blog/race-gender-and-popular-culture/202408/falling-out-of-the-coconut-tree-what-the-popular-kamala

Comments URL: https://news.ycombinator.com/item?id=47304586

Points: 1

# Comments: 1

Hi HN,

I’ve been working on a project called OpenVerb, which explores an architectural idea for AI systems: separating reasoning from execution.

Most AI agent frameworks today focus on improving reasoning loops, planning, and orchestration (LangChain, LangGraph, etc.). But once an agent decides to perform an action, execution usually becomes a direct tool call, script, or API invocation.

That approach works, but it also creates some issues: • custom glue code for every integration • inconsistent action schemas • limited determinism in execution • difficult auditing and policy enforcement

OpenVerb experiments with treating actions as a protocol layer, not just function calls.

Instead of arbitrary tool calls, systems define structured verbs that describe: • the action being performed • required inputs • expected outputs • execution policies • audit information

Conceptually the architecture looks like this:

AI Model / Agent Framework ↓ Reasoning Layer ↓ OpenVerb (Action Protocol) ↓ System Execution

The idea is that agent frameworks control how the AI thinks, while OpenVerb standardizes how actions are executed.

Some existing projects touch related areas: • Model Context Protocol (MCP) – tool and data discovery for AI systems • LangGraph – deterministic reasoning loops for agents • PydanticAI – structured schemas for agent outputs

OpenVerb is trying to explore something slightly different: a universal grammar for deterministic execution that could work across domains (software systems, spatial systems, robotics, etc.).

Still early and experimental, but I’d love feedback from people thinking about agent architecture or execution reliability.

Curious if others have explored similar ideas or if there are related systems I should look at.

Comments URL: https://news.ycombinator.com/item?id=47304573

Points: 1

# Comments: 0

Article URL: https://github.com/followtayeeb/llm-costs

Comments URL: https://news.ycombinator.com/item?id=47304545

Points: 1

# Comments: 0

We build RobotActions, a cloud device farm for Android/iOS testing. We just shipped a Chat AI Agent that lives inside the live device session.

What it does:

- During a session, you can ask "What's the XPath for this button?" and get a ready-to-use locator from the current screen - Ask "Write an Appium test for this flow" → get test code generated from the live accessibility tree - Type "tap the login button" in natural language → it executes on the real device - Ask "Why is my test failing on this element?" → gets context from both vision and the accessibility snapshot

The agent uses a combination of screenshot vision and the device's live accessibility tree. The key insight is that most mobile test failures are locator issues or UI state issues — and an agent with full context of what's on screen right now can solve those immediately, without the engineer leaving the session to use a separate inspector tool.

Technical bits: - Accessibility tree is captured per-frame during the session - Agent has both visual context (screenshot) and structured context (a11y tree) simultaneously - Supports Android (UIAutomator2/XPath/UISelector) and iOS (XCUITest/Appium) - Session context is also exposed via API for CI/CD post-failure reports

Happy to discuss the architecture, especially the tradeoffs between using vision alone vs. vision + a11y tree for locator generation.

Comments URL: https://news.ycombinator.com/item?id=47304542

Points: 1

# Comments: 0

LLM coding agents (Claude Code, Codex, etc.) have structural weaknesses: blind retry loops, volatile learning, silent spec drift, and gate gaming. These aren't bugs in specific models — they're properties of goal-optimizing systems.

ANDON applies Toyota Production System principles to address this: - Jidoka (autonomation): auto-detect failures and block forward-progress commands (git push, deploy) - Kaizen (continuous improvement): force Five Whys root cause analysis, then standardize prevention rules - Meta-ANDON: detect when the agent is stuck in a whack-a-mole debugging loop Install: pip install andon-for-llm-agents Works with any agent that supports hooks/callbacks. Apache-2.0 licensed.

Comments URL: https://news.ycombinator.com/item?id=47304539

Points: 1

# Comments: 0