igorpcosta

There's not a lot of opportunities in this space yet. This is the closest we can get to High degree solver kinda of problem.

There are only 3 companies doing this to date: Google, Sakana AI and Autohand AI.

$2B dollar in funding, yet unable to solve a problem, but getting ready for another series of funding Round to keep the machine afloat.

This made me laugh https://www.forbes.com/sites/annatong/2026/03/05/cursor-goes...

If that's what winning looks like AI Summer is coming!

tell me about it, it's so frustrating

I think it's extremely value what Thomas is doing with Checkpoints, I have this built in on autohand code cli that allows you to share with others your session /share https://autohand.link/s/ah-23wbbq4v2r5vt86684hg-1768476838

You can see the turns, tools called, outputs, changes. Similar to what he's trying to achieve.

Cool concept, however relying on Actions are really not guaranteed execution all the time.

I'd love to have you as a top user on autohand [.] ai/cli and interested in your experience with us. We're a bootstrap ai lab.

I'd love to have you as a top user on autohand [.] ai/cli and interested in your experience with us.

Hey, author here.

My view is that software engineering is splitting into two paths. One is deep craftsmanship. The other is guided operators who know how to steer tools, systems, and LLMs toward the right result.

I am keen to hear where others agree or disagree, and how you see this playing out in real teams.

Open source repo: https://github.com/autohandai/code-cli

Most coding CLIs I've seen lock you into one provider or requires you to bypass by changing BASE_URL and has a lot of conflict. That works fine if you're committed to one vendor for coding cli harness, but it breaks down when you want to run local models, test different providers, or avoid API costs entirely. So we tried a different approach. Instead of hardcoding a provider, Autohand code lets you swap between OpenRouter, Anthropic, OpenAI, Ollama, llama.cpp, and MLX from the same codebase. Switch models mid-conversation if you want. High level, the design optimizes for three things:

Machine orchestration: stateless execution, structured outputs, designed for CI/CD and batch runs, not just interactive use Auto mode: autohand -p "fix the tests" --yes --auto-commit runs the full task without prompts. Three permission levels plus granular command whitelist/blacklist Skills system: modular instruction packages that activate on demand. Run --auto-skill and it generates skills tailored to your project

One thing that's been surprisingly useful: because it's provider-agnostic, you can prototype with a fast cheap model, then swap to something heavier for the actual run. No code changes, just config. It's TypeScript + Bun, 40+ tools (file ops, full git, semantic search, multi-file edits), sessions persist and resume.

Try autohand.ai/cli/ most of these problems you’ve mentioned here are solved already

Hi HN, Igor here, founder of Autohand.

Today we are sharing a deep technical guide on how we built Git Flow Automation in our Evolve platform. It solves a real engineering problem: how to run hundreds of agent tasks in parallel across a codebase without conflicts, without slowing down, and without breaking CI.

Rather than simple scripts or CI tricks, we use Git worktrees to give each agent its own isolated branch and working directory. This lets each agent run tests, detect conflicts, and even roll back individual tasks. We also built conflict detection, AI assisted resolution, and merge strategies that keep history clean and safe for teams.

Here is my take based on 20+ years of experience in Devtools and dealing with lots of large code bases.

The guide walks through:

Why sequential execution fails at scale

How parallel worktree orchestration works

Task lifecycle and dependency ordering

Conflict detection and automatic resolution

Testing per task and rollback controls

Merge strategies and commit hygiene

Safety limits and observability tooling

This is a technical system design share, not a product announcement. I would love feedback from builders and maintainers who work on large codebases or autonomous tooling.

Super interested in read different ideas.

Read the guide here:

https://autohand.ai/docs/guides/git-flow-automation

Hi HN, founder here of Autohand AI

A few weeks ago I shared some thoughts on intent weaving in AI coding agents. Today I would like to invite the community to our private beta of Evolve.

Evolve is an AI software engineer designed for long horizon workloads. It uses reinforcement learning style feedback loops to learn from past mistakes, improve future decisions, and reduce repeated human intervention when working on real codebases.

We are early and actively looking for feedback from people who deal with complex systems, legacy code, or long running engineering tasks.

Private beta: https://autohand.ai/products/evolve/

Happy to answer questions and hear what does not make sense.

Very interesting research on this, keen to colab with you folks, I've been building a few experiments for old GTX GPUs to extend lifetime of them with matching performance of tokens for Smol, igor [] autohand.ai let's chat.

As we build the capability, our intention is to give the tools for companies to compete with FAANGs, even if they don’t have the engineering excellence to do.

Today we are warning the entire community, stop using companies providing APIs for coding tasks on your company source code.

I’d love to have different opinions here and create a better version of the future together.

Addy Osmani also wrote today about this https://addyo.substack.com/p/conductors-to-orchestrators-the...

I wrote this before I left GitHub and how I’d like to see Microsoft invest in this area, but it was too futuristic, considering how the market has evolved today, I decided to make this public.

If you want to know more about what I’m doing next www.autohand.ai

Wow, OCR is now basically a general domain. I remember when I spent like a year trying to create one for receipts. Took me 6 months of data curation to prepare.

Nice job, the scores are superb.

That's the intent, more people with different background can help and build different solutions for this problem.

That's a real worry that motivated us to build Autohand. We're going to build a platform to empower companies to have the same level of accuracy and faithfulness without having to worry about that. There are a vast number of data and code that haven't been exposed yet to these model providers.

Very interest @asdev, I love this kind of simple straight to the point UI, Would you love to join forces on https://github.com/autohandai/commander

It's not fully functional now as we're building this in the open. Fully supports the 3 major ones, Codex, Claude, Gemini. But I'm also working on a version that doesn't use git worktree, some enterprise folks don't like it. So we're solving this by adding an isolated environment, similar to what others are trying to establish.

that's very cool, haven't thought about that, how do you score the policy if in case they conflict or override previous directives?

This is super cool, Specially for VRs and Augmented Ray-bans. I still prefer a mechanical keyboard to type stuff.

Hey HN community, I’m Igor, co-founder of Autohand.ai in New Zealand.

We’re building an open stack that lets AI coding agents deliver work with the discipline senior engineers expect. Our latest write-up, “Intent Weaving for AI Coding Agents,” breaks down how we encode strategy, policy, and telemetry into machine-executable intent, plus an honest inventory of where current agents fail (reasoning, repo awareness, testing, etc.).

Highlights: - Mission compiler that turns business objectives into guardrail-rich plans for agents. - Knowledge graph + policy DSL so automation stays inside governance envelopes. - Pain-points matrix from real deployments; new benchmarks that punish regressions, not just pass unit tests. - Open-source pieces as we release them; Commander is already MIT-licensed.

We’d love feedback from folks shipping agentic workflows or wrestling with AI codegen drift. Where should we push harder? What failure modes have we missed?

Link to our manifesto: https://autohand.ai/manifesto

Thanks for reading, and be kind. Creating a new category means stretching before the skills are perfect.

That's really sad news.

That's sad news, just got the attention now.