Last week a stop gate refused to let my agent finish, twice in one session. Once on six regressions it had just written. Once on a placeholder it left in its own notes. The "done" message was equally confident both times.
If an agent has done this to you on a real codebase, I want the story, not the fix.
3
My own agent reported "done" on a sync change. The tests it picked passed. Three months of cloud notes were gone behind a swallowed error and a 200. I found it by counting rows in prod.
That's mine. What's yours, and what did it cost?
1
6
Meta just released Muse Ember, a smaller model built for high-volume agent workflows like tool calls, coding subtasks, and document processing.
Meta claims it can match Claude Haiku 5.5’s performance at roughly a third of the cost.
I’m interested in this direction. As we build more complex agents, model selection becomes an architectural decision. Having a powerful model handle planning while smaller models execute routine tasks could make these systems significantly more economical.
I’d like to see how Ember performs in real-world agent workflows.
17
Anthropic just dropped Claude Fable 5.2 and Mythos 5.2. Same underlying model, different safeguards and access. Up to 30% faster output.
I like Anthropic’s approach to using Fable as an advisor to smaller agents. Save the heavy compute for problems that actually need it.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
23
Been testing Grok Bot on my dev work this week. Nowhere near the setup Lauren @poteto has running, but I am thoroughly impressed.
I already hit my weekly usage limit. Might be time to bump up my subscription.
2
56
I just folded the mechanical parts of @poteto's pstack into Ratchet: a rule-to-enforcer table so prose-only rules get promoted to checks, "inconclusive" as a verdict that never passes, and a pickup skill that verifies a handoff instead of trusting it.
github.com/baisethomas/Ratch…
9
Anthropic, OpenAI, Hugging Face and Alibaba just moved four agent projects into open governance: Agent Skills, Open Responses, OpenEnv and OpenSandbox.
I like this. If I swap models or tools next month, I shouldn't have to rebuild the workflows around them.
1
12
Anthropic dropped Claude Duet today. Meta has Muse, OpenAI has Dots, now Claude has an agent that comes to you first.
I live with a proactive agent every day. The autonomy part is easy. Judgment is the whole game: knowing what needs you, what can wait, and when to handle it.
20
My son Micah has been making Minecraft add-ons, so we built him a site to share them.
Roller coasters, water parks, extra TNT. All free for Minecraft Education.
He also has a “Buy me a cookie” button now lol.
Pretty cool building this with him.
mi-craft.com
19
Anthropic found its employee-directed agents worked 6x as many hours as staff by August. But Claude sessions spent ~80% of their time waiting for human input.
For builders, that’s a clue: better triage and review may matter more than another bump in model speed.
4
OpenAI is bringing GPT-6 into Atlassian, grounded in Jira, Confluence and the context around how teams actually work.
This is where enterprise agents get interesting.
The model matters, but access to the right context at the right time might matter more.
14
Mistral just dropped Large 4: a 1T-parameter open-weight multimodal model with 49B active parameters and a 1M context window.
Open models aren’t just getting better. They’re getting harder to ignore.
For builders, owning more of the stack is becoming a real option.
12
Anthropic is spending $100M to train 10,000 Frontier Deployed Engineers.
That’s a pretty strong signal.
The AI skill gap isn’t just building models. It’s knowing how to take them into a company, connect the messy pieces and make them actually useful.
7
OpenAI says its agents accidentally sent 53 ChatGPT user images to third-party sites.
We’re getting good at making agents capable.
I’m not convinced we’re equally good at controlling what they can access.
Permissions and observability might be the next big agent problem.
1
25
Microsoft might finally be catching up.
Copilot now has its own memory, computer and workspace so it can keep working after you leave.
This feels a lot closer to where agents have been heading.
Now Microsoft has to make it useful.
12
I built Ratchet, a playbook for coding agents.
The model's ceiling is fixed. The system's isn't. Every failure that gets past you adds one permanent fix: a hook, a test, a rule.
Any model. Now a Claude Code plugin. Open source, MIT.
github.com/baisethomas/Ratch…
1
1
70


