Blurbs

If you want a compute and token efficient 'agentic systems' your goal should be to use your non deterministic coding tools to replace as many parts of your non deterministic system as possible with deterministic code.

Just not all of them.

Start off with something wildly inefficient then invest in evolution towards efficiency.

I've been pondering what multi-player agent workflows look like in practice. As in multiple humans and multiple agents collaborating together. How do people work together at people speed while agents work at agent speeds.

Even the folks at GitHub acknowledge that the future of collaborative agent driven coding isn't going to be issue queues and pull requests. Not the way that teams have been used to.

This prototype looks promising: (Demo around the 7 minute mark)

I'm curious if anyone is building agent harnesses with idempotence via some mechanism where outbox and/or saga patterns are spun up on the fly and the critical agents execution goals are locked in until completion.

Those are big words, what does that mean Andre?

Imagine you specify your goal, and your agents come up with a plan. Instead of just spinning up a todo list they come up with real contracts for themselves where all obligations of all contracts are completed or none of it is considered complete.

I'm sure this is done in bespoke 'deterministic' agent flows, but I'm wondering if anyone is baking it into a general purpose harness.

I'm thinking out loud here, but I imagine this kind of thing could be composed with existing harnesses, skills and tools. That said, I'm wondering if a standard is emerging somewhere.

The naive or knee-jerk "can't teach an old dog new tricks" mentality might presume to close the door on 'expensive' experienced talent.

But capital W Wisdom is critical for maximum AI leverage. That Wisdom is gained over time. Knowing the right questions to ask and giving informed direction to people or agents yields better results that compound quicker.

I might ask the question of whether AI is accelerating or decelerating aging in.

Here too the naive "ai is as good as any junior and doesn't complain" mentality might presume to close the door on 'cheap' but inexperienced talent.

But once that Wisdom does actually age out, who in any business is left with institutional memory, domain knowledge, or real judgment?

I've been testing Gemma4 models running locally today. Believe the hype. They're fast and the smaller 4B model is doing a great job with NLP and NER tasks.

• e2b is super fast but overly concise
• e2b misses more secondary entities/typed mentions (lower recall).
• e4b gives richer extraction (more entities, more nuanced categories), but with more noise/normalization quirks - and takes about twice as long to process.

The results for the task were as good or better thang GPT5-mini. I'm not sure if OpenAI was having issues today, but performance was much faster than GPT5-mini over the network today.

Mid stage enshitification of GitHub and GitHub Copilot: Ads in pull requests. They walked it back, but the advertisements were an example of what I posted about yesterday: Abdication of responsibility to exercise good judgement.

To the Github Copilot team's credit they explicitly called out 'poor judgement' as the core issue and took some responsibility after the fact.

If you take their stated goal at face value the team was trying to surface 'helpful tips' to users. How the defined a 'tip' and where they chose to surface them both lacked appropriate judgement. To a point a friend recently commented, they met a specification (ads in pull requests) but not the requirements (helpful tips for users).

Zack Manson connected the dots to Doctorow for me with this quote on enshitification:

Cory Doctorow