nileback

Co-founded @voidleapcom. Better models will not fix bad tools. Building Voidleap Code, an agentic IDE that kills the black box. Bootstrapped in Sweden.

[nileback @nileback]

Another coding agent wrapper is a hard sell. We built an agentic IDE instead.

Models keep improving. What runs them has to keep up. I need to see what my agents did and how, or I am just guessing.

Free. Mac and Windows. Download at https://t.co/u5b0v0w1eQ

[nileback @nileback]

Does @voidleapcom make you a better engineer?

We built Voidleap Code: an agentic IDE with our own agent harness, so engineers control how they build software with AI agents.

Our AI Gothenburg pitch. Watch. Try. Judge for yourself.

https://t.co/8661Vkkmtc

[nileback @nileback]

@Layton_Gott Change harness. Use subagents as tools (not roles). Appoint simpler models for subagents.

Seen it reduce token cost down to 25%. Not by 25%. Down to 25%.

Still no where close to subscription-level subsidisation though.

[nileback @nileback]

How to explain AI coding harnesses? Move your hands like an accordion while explaining the screen.

Pitched for 65+ investors last week. With a two-minute pitch I summarised the why and what we spent thousands of hours to build.

In-depth talks afterwards gave solid insights!

[Voidleap Code @voidleapcom]

"I've got too many ongoing threads to manage across my projects!"

After feedback from the developer community, Voidleap Code just added a Kanban board.

Let us know what more can help you build software better.

[Voidleap Code @voidleapcom]

Feature drop: Schedule your agent runs

You know what the agent should do next, but you don’t want to send the prompt yet.

Pick a time. Queue it. Or repeat it on a schedule.

[nileback @nileback]

What's not evaluated in the report is how your agentic setup and workflows affect quality or cost. It make a huge difference in quality and token usage.

[nileback @nileback]

AI fact checking: Only AI Labs can do model-harness optimization well.

Verdict: Wrong.

"The Harness Tax" report shows that independent harnesses have better token efficiency than the AI Labs own harnesses.

https://t.co/scOZuXOwB8

[nileback @nileback]

Token cost varied largely, but correctness was comparable. So essentially, you pay a "token tax" depending on which harness you use.

For subscription users, this translates to more or less output per weekly limit.

For API customers this translates to actual cost.

[nileback @nileback]

@brannonhogue Did you trust the model to act without guardrails?

[nileback @nileback]

@shownotover During peak hours as well?

[nileback @nileback]

@StefanoGPT Did you notice the improvement during or outside peak hours? Cache hit-rate have been low during peak hours, and not great off either. Server-side issue.

[nileback @nileback]

@CHESSWIZARD5 @callebtc OpenAI seems to have a server-side cache write delay, causing substantial cache hit rate reduction. That's where your usage is spent.

[nileback @nileback]

@AnkanXplorer @Im_IrushiK Seem to be OpenAI server-side cache issues

[nileback @nileback]

@PovilasKorop OpenAI seems to have a cache issue at the moment. It may be warranted to retry after they fixed that if you want to compare cost/usage.

[nileback @nileback]

@coryparrry Seems like they have a cache write issue. Written cache not available for up to 40 seconds, so many of your tool calls draw full price. I guess that increases load on inference engine, reducing availability.

[nileback @nileback]

@Im_IrushiK No they have a cache write issue. Written cache not available for up to 40 seconds, so many of your tool calls draw full price. I guess that increases load on inference engine, reducing availability.

[nileback @nileback]

@nomdk1 During peak hours? The root cause of recent ChatGPT subscription issues might be slow cache writes, specifically during peak hours.

[nileback @nileback]

Cache write delays cost you 3-4x more usage!

For ChatGPT subscription, everyone been reporting usage issues and burning weekly limits quicker than normal.

This guy building a harness might have found the root cause. It's server side, not Codex.

https://t.co/BKig8F61hF

[nileback @nileback]

@Voxyz_ai Is there any pre-compaction actions done in Codex? Like trimming tool calls before compacting.