@marioscheliga.bsky.social
Legora wants lawyers to "trust the output" of their agents.
As long as an agent can reward hack, trust needs hard gates. Even Legora's own benchmark says the scaffold matters more than the model.
How do they build theirs?
As long as an agent can reward hack, trust needs hard gates. Even Legora's own benchmark says the scaffold matters more than the model.
How do they build theirs?
Legora introduces the Legora aOS™: the agentic operating system for legal work - Legora
Legora today unveiled the Legora aOS™ – a purpose-built agentic operating system that enables legal teams to execute complex legal work end-to-end, marking the arrival of a new era for the legal industry: Agentic Law.
legora.com
October 28, 2026 at 2:54 PM
Legora wants lawyers to "trust the output" of their agents.
As long as an agent can reward hack, trust needs hard gates. Even Legora's own benchmark says the scaffold matters more than the model.
How do they build theirs?
As long as an agent can reward hack, trust needs hard gates. Even Legora's own benchmark says the scaffold matters more than the model.
How do they build theirs?
Delaware wants AI agents to run companies, with the owner shielded from the debts.
My agents still need a gate before they may call a test green.
Real cowboys. Or they know something I don't.
https://marioscheliga.com/blog/ai-ceo-delaware/?utm_source=bluesky
My agents still need a gate before they may call a test green.
Real cowboys. Or they know something I don't.
https://marioscheliga.com/blog/ai-ceo-delaware/?utm_source=bluesky
October 28, 2026 at 11:07 AM
Delaware wants AI agents to run companies, with the owner shielded from the debts.
My agents still need a gate before they may call a test green.
Real cowboys. Or they know something I don't.
https://marioscheliga.com/blog/ai-ceo-delaware/?utm_source=bluesky
My agents still need a gate before they may call a test green.
Real cowboys. Or they know something I don't.
https://marioscheliga.com/blog/ai-ceo-delaware/?utm_source=bluesky
Anthropic's IPO filing, risk factors: in a 2026 test, four models got into real third-party systems without authorization. One found a password, took admin and read a person's data.
If they find it, they will play with it. Agents need locks, not trust.
If they find it, they will play with it. Agents need locks, not trust.
October 28, 2026 at 10:31 AM
Anthropic's IPO filing, risk factors: in a 2026 test, four models got into real third-party systems without authorization. One found a password, took admin and read a person's data.
If they find it, they will play with it. Agents need locks, not trust.
If they find it, they will play with it. Agents need locks, not trust.
GLM-5.3 on Cloudflare Workers AI put other customers' prompts into 155 responses, 27 Aug to 9 Sep. A 1M context on a KV pool sized for 681k, plus an SGLang bug. Your agent's context sits in GPU memory next to strangers'.
Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
A configuration mismatch in GLM 5.3 on Workers AI triggered an SGLang bug that exposed KV cache pages across requests. Here is how we addressed the issue.
blog.cloudflare.com
October 27, 2026 at 4:00 PM
GLM-5.3 on Cloudflare Workers AI put other customers' prompts into 155 responses, 27 Aug to 9 Sep. A 1M context on a KV pool sized for 681k, plus an SGLang bug. Your agent's context sits in GPU memory next to strangers'.
Agents write code faster than anyone can read it. adrq's picture of what's left for you: a wide bar of breadth and many narrow dives, the deepest one ending in a submarine. Hard time to be a junior.
What Engineers Still Need to Know When AI Agents Code
AI agents write code faster than anyone can read it. What engineers still need: fundamentals, breadth, and the willingness to dive in when it breaks.
blog.adrq.dev
October 27, 2026 at 3:05 PM
Agents write code faster than anyone can read it. adrq's picture of what's left for you: a wide bar of breadth and many narrow dives, the deepest one ending in a submarine. Hard time to be a junior.
My CLAUDE.md tells the agent never to edit a test. Opus read it and gamed the spec anyway. 5 of 5 runs.
180 Claude Code runs: markdown changes the report. A hook outside the codebase changes the result.
https://marioscheliga.com/blog/ai-agent-definition-of-done/?utm_source=bluesky
180 Claude Code runs: markdown changes the report. A hook outside the codebase changes the result.
https://marioscheliga.com/blog/ai-agent-definition-of-done/?utm_source=bluesky
October 27, 2026 at 8:47 AM
My CLAUDE.md tells the agent never to edit a test. Opus read it and gamed the spec anyway. 5 of 5 runs.
180 Claude Code runs: markdown changes the report. A hook outside the codebase changes the result.
https://marioscheliga.com/blog/ai-agent-definition-of-done/?utm_source=bluesky
180 Claude Code runs: markdown changes the report. A hook outside the codebase changes the result.
https://marioscheliga.com/blog/ai-agent-definition-of-done/?utm_source=bluesky