Leo Chen

Building Orkas — open-source, local-first multi-agent desktop (2.1k ⭐).
I write about agent architecture, where AI is heading, and the messy parts of shipping.

[Leo Chen @leochenpm]

Agent frameworks make you write orchestration code. Cloud platforms make you hand over your data.

Orkas is neither — an open-source, local-first multi-agent desktop app where you describe the goal in chat and a super-powered commander leads a team of specialist agents that do the work for you. No flowcharts. Nothing leaves your disk.

https://t.co/bjHJYANt0d

[Leo Chen @leochenpm]

1、There’s a huge gap between what AI can do and what many users can actually accomplish with it. Closing that gap is a major opportunity for AI products.

2、Better models make it easier to use AI well. They also expand the range and complexity of tasks AI can tackle. The gap keeps moving as the ceiling rises.

3、In mobile apps, onboarding mostly uses UI to teach fixed features. It accounts for a small part of the product’s value. Once users learn the product, they rarely need that guidance again.

4、AI products need agents that continuously help users get more from model intelligence. Capabilities keep improving, but remain jagged. Until we reach superintelligence, bridging that gap will remain a core source of product value.

5、That’s the role of Orkas’s Commander agent. Whether you’re learning the product, working through uncertainty, or figuring out how to keep making progress toward a goal, you can ask it.

[Felix Rieseberg @felixrieseberg]

Hi, I'm Felix. I'm an eng lead working on Claude. I love watching people take flight and do things they never thought possible.

We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial:

https://t.co/vNPPs5FWm8

[Leo Chen @leochenpm]

@Verdana86 Yes, this is crucial. We need a mechanism to detect whether an agent is producing hallucinated conclusions.

[Leo Chen @leochenpm]

Two of our agents read one stale field on our analytics dashboard: "can't be linked." (It could be.) Both quietly dropped an entire user-journey analysis.

A person would have asked: "is this still right?"

That's the core problem when people and agents share a project.

Chat apps were built around human limits. We can't hold much in our heads, so we sync constantly. We meet in person to read tone and build trust.

Agents are the reverse: huge bandwidth, end to end from PRD to ship, no trust to earn. But they:
• need clean context
• misattribute causes over long horizons
• adapt badly to changes they never saw in training

So in Orkas we kept what still works:
• a human owns and reviews every task
• people and agents read the same goals, rules and memory

And redesigned the rest:
• hand off through files
• keep the messy middle out of other people's agents
• let teammates share their personal agents

It's early, and models will absorb a lot of this.

Full write-up:
https://t.co/yf21gjHZmo

[Leo Chen @leochenpm]

@Verdana86 When a human reviewer noticed the analysis was incomplete, we asked the agent to investigate why. The agent traced the issue to a specific field, and we ran an A/B test to confirm that the field was indeed the cause.

[Leo Chen @leochenpm]

@jqdsouza Cross-harness memory is useful, but how do we ensure these skills remain useful as the underlying models powering Claude Code and Codex keep improving? As new models come out, many skills offer diminishing benefits—and some can even hurt performance.

[Leo Chen @leochenpm]

DeepSeek just open-sourced DeepInfer, the engine behind V4.1-Flash (155T tokens/day per cluster).

By my math from the code, each token of context costs this much KV cache:

• Llama-3.3-70B: ~164 KB
• DeepSeek V3: ~35 KB
• DeepSeek V4.1-Flash: ~0.89 KB

How they got there:

1. Only 4 of 40 layers write KV (layers 2, 8, 14, 20). 34 layers reuse theirs, and the whole decoder half shares layer 20's cache, YOCO-style. That alone is ~10x.
2. FP4 storage: 288 B per entry vs 528 B at FP8.
3. 2:1 pooling of adjacent tokens in three of those layers.
4. A small FP4 index on top.

Long context still works: every layer reads the last 128 tokens exactly, and older context goes through a sparse index that picks the top 512 entries.

The payoff: a 100K-token session's KV is ~89 MB instead of ~16 GB. That's what makes keeping agent caches on SSD for 3 days practical.

Caveats: my estimate from the config and code, not an official figure (~1.5 KB at FP8). FP4 KV needs SM100 or Ascend 950. Quality numbers are self-reported.

[Leo Chen @leochenpm]

@ryolu_ @brian_lovin WeChat has the strongest ecosystem and moat, yet acts like it has nothing to prove.

[Leo Chen @leochenpm]

@vitverb Yes, Orkas has a to-do board where both users and agents can see unfinished tasks. These tasks also stay in the Commander agent’s context, so it can remind users about outstanding work at appropriate moments.

[Leo Chen @leochenpm]

Build your small business. Move what matters forward.

Orkas 1.9.0 is here!

1. Stay focused on what matters, one step at a time
Turn the next steps for product launches, customer feedback and other important work into to-dos. Unfinished work and task results you haven't reviewed appear together, so you can see what needs your attention. Whenever you return, pick up where you left off.

2. Give your business an online presence and reach more people
Ask Orkas to create a product reservation page, a page introducing your services or an event sign-up form. Publish it and share it with your communities and on social media to help more people discover your products and services. Customers can open the link to browse and fill in forms, while you check page views and submissions to follow up on their needs.

3. Work with your partners, even on a Free account
Invite partners to a shared project to share materials, clarify requirements and assign to-dos. Everyone can see who's responsible for what and how the work is progressing. From launch preparation to follow-up, keep things moving together.

website: https://t.co/tDNa5aIPuF

[Leo Chen @leochenpm]

@aiwithadb https://t.co/1G6XcQLdPn

[Leo Chen @leochenpm]

@Joi2James Orkas (https://t.co/zQlxFy5nZW): Direct a team of AI specialists to deliver the work. For indie builders, one-person businesses, and small teams

[Leo Chen @leochenpm]

@TChudiwal All sorts of similar errors.

[Leo Chen @leochenpm]

@maisonnatmax We’ve run into this many times ourselves, so we’ve focused more on curating high-quality context.

[Leo Chen @leochenpm]

Interesting

[Felix Rieseberg @felixrieseberg]

Hi, I'm Felix. I'm an eng lead working on Claude. I love watching people take flight and do things they never thought possible.

We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial:

https://t.co/vNPPs5FWm8

[Leo Chen @leochenpm]

@uncle_moon_x 很少看到明星创业者去做非常非共识的事情，更多的是聚集在有市场热度的领域。

[Leo Chen @leochenpm]

LeCun's AMI Labs open-sourced H-JEPA: world models stacked by timescale. The top plans in 20-step jumps, handing subgoals down.

Maze success vs a flat JEPA: 18% → 73%, on less planner compute.

Unprompted, the top kept position, mostly dropped leg pose.

https://t.co/iOVqli7PUC

[Leo Chen @leochenpm]

@uncle_moon_x 现在这种氛围有点畸形

[Leo Chen @leochenpm]

@Ashyedra Exactly. Handing off simpler execution tasks to Haiku can cut costs significantly compared to before.

[Leo Chen @leochenpm]

Haiku 5.5 just launched at $0.10/M input, same as GPT-6 Luna and below DeepSeek V4.1 Flash and GLM-5.3 Flash.

But per-token price ≠ cost per task: at max effort Haiku outputs ~162K tokens per eval task vs Luna's ~50K, and bills 5x past 100K context.

What's actually in your prod stack?

[Leo Chen @leochenpm]

@mylifcc 不管多底层的代码，只要有数据模型都能学得会，现在模型进步比较慢方面是，在业务约束下的架构趋势，仓库的长期可维护性等等，本质也是训练数据不太好构建