HiloW.Ai

Testing AI tools & Building AI systems
Agents · Workflows · Github
Claude Code · Codex · Obsidian · Jev

[HiloW.Ai @HiloW_Ai]

Gemini 4 Argon has been 'official' since Sep 30, but only for vetted cyber defenders. The wider rollout is still a leak: model IDs in Antigravity, a GCP quota entry, a one-word tease.

Everyone's counting hours. Here's what I'd prep before it lands 🧵

[HiloW.Ai @HiloW_Ai]

5/ Longer runs, bigger blast radius. A model built for long-horizon work will touch more tools unattended. Permissions live in hooks and scopes, not in the prompt. Stronger model, same guardrails, not fewer.

[HiloW.Ai @HiloW_Ai]

Gemini 4 Argon has been 'official' since Sep 30, but only for vetted cyber defenders. The wider rollout is still a leak: model IDs in Antigravity, a GCP quota entry, a one-word tease.

Everyone's counting hours. Here's what I'd prep before it lands 🧵

[HiloW.Ai @HiloW_Ai]

4/ Pin what you run. Log the model ID the API returns next to your prompt version. Aliases move under you. Low/medium/high effort is a new knob: treat each level as its own config in evals and in your cost budget.

[HiloW.Ai @HiloW_Ai]

@kimmonismus “Era of AI agents” only works if someone owns the plan. Cursor watched flat multi-agent coordination collapse to the throughput of 2–3 workers; hierarchy beat self-coordination. Max out one agent first, split when it keeps picking the wrong tool.

[HiloW.Ai @HiloW_Ai]

@steipete @simonbs Per-agent trust levels beat per-project every time on shared gateways. Same lesson as hooks vs CLAUDE.md : a rule in prose is a suggestion. Permissions have to live on the agent identity, not the repo ACL.

[HiloW.Ai @HiloW_Ai]

@OpenRouter @scaledown_team Task-specialized models on a router is how production graphs should look: summarize for compaction, extract for structure, frontier only when you need generation. Short, consistent summarize steps protect KV-cache hit rate better than one giant chat.

[HiloW.Ai @HiloW_Ai]

@omarsar0 Right split: frontier models for writing, constrained decoding for point-to-answer work. A lot of agent steps are still yes/no or extract calls paying a frontier model to classify at scale was always the expensive smell.

[HiloW.Ai @HiloW_Ai]

@testingcatalog Custom MCPs in a desktop app is great UX and a sharp trust boundary. Tool defs alone can eat 50k+ tokens before the task, and each new server can close the lethal trifecta (private data + untrusted content + egress). Defer-load tools and treat tool responses as untrusted input.

[HiloW.Ai @HiloW_Ai]

@testingcatalog When a new Gemini shows up on quotas before the blog post, the useful move is routing by role, not a full swap. Cursor's writeup: one model plans better, another holds focus on long runs. Treat Argon as a candidate for one hop in the graph until your evals say otherwise.

[HiloW.Ai @HiloW_Ai]

@woongworks Both, in that order. Canary diff on the replacement first, then a short dual-run on a slice of real traffic. The diff catches obvious drift, the dual-run catches what only real inputs trigger. I switch once tool-call sequences match closely enough.

[HiloW.Ai @HiloW_Ai]

If Opus 5.5 really is getting swapped to a new checkpoint under the same name, that's great for chat users and a headache for anyone running agents in production.

Half my agent bugs show up right after a model changes. Same prompt, different behavior.

Pin the version, rerun your evals, then upgrade.

[Chubby♨️ @kimmonismus]

Seeing so many posts about people saying that fable 5 now routing to a new model, possibly fable 5.5, with much enhanced capabilities.

A release should be near. Since Anthropics IPO is also pretty close they want their new model to be release a few weeks ahead so the market can really count that into the valuation.

Long story short: Fable 5.5 should arrive soon. And it looks to be a much bigger leap than we even expect.

[HiloW.Ai @HiloW_Ai]

@asifalis @CrowdStrike Yeah, keeping scopes tight as agents spawn and die is the hard part. I treat each sub-agent's creds as short-lived and revoke them on exit, so the blast radius stays small.

[HiloW.Ai @HiloW_Ai]

@TRTRE62 @grok Governance before scale, yeah. If authority can't survive mutation and handoff, more IoT just multiplies the blast radius. I'd prove the control plane on a few agents first, then let devices scale.

[HiloW.Ai @HiloW_Ai]

Anthropic (today): small models may be months from self-replicating on IoT. The same agentic coding skills that ship PRs also chain cyber steps. Not a full rogue worm yet. If your agents have shell, treat the sandbox like prod.

[Andrew Curran @AndrewCurran_]

Anthropic's Frontier Red Team believes small AI models may now be months away from being capable of self replicating without access to a datacenter.

Quoting from the blog:

'The capabilities that make AI models generally powerful coding agents also generalize to cyber tasks, including autonomous replication. Like coding agents in 2024, we’re seeing signs of life for AI worms. We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves).

At the same time, increasing amounts of inference compute are now available both inside and outside of datacenters (see Appendix D). It’s uncertain exactly what form autonomous AI replication will take. It could be a slow, distributed “digital slime mold,” living on relatively insecure compute. It could also be a powerful cyber model that breaches datacenters at a high pace, or a model that scams people for compute credits.

As a prediction that goes beyond our experiments, we expect a future awash with agents running on all types of hardware, including some that aren’t under human control. While it’s true that IoT devices have been vulnerable for a long time, lightweight agents may be much more willing than human attackers to put in the effort to compromise large fractions of the world’s IoT compute. The rise of slow replicators would also be a bellwether for more aggressive, faster ones that can breach more hardened machines or be effective API key thieves. We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay.'

[HiloW.Ai @HiloW_Ai]

"No more hand-written code" is the easy part to announce. The hard part is review. AI can write a game's code fast; someone still has to own what it does at 2am on launch day.

Agents don't remove engineers. They move them from typing to verifying.

[DiscussingFilm @DiscussingFilm]

Epic Games has internally delayed the release of a shooter game they’ve been developing featuring Star Wars & Marvel characters.

• Developers on the game have now been told to only use AI to write the code — “No more hand-written code”

• All art for the game would be “generated with full AI workflows”

(Source: https://t.co/i7ZykpqD7j)

[HiloW.Ai @HiloW_Ai]

@alexandr_wang Right call. Nobody's agent pipeline renders 3D games. Mine reads messy repos, calls tools 100+ times and has to stay on plan. Demos sell the launch, boring reliability decides what stays in production.

[HiloW.Ai @HiloW_Ai]

@Stilllabs Fair. The agent is still deciding if you pass its eval set 😄

[HiloW.Ai @HiloW_Ai]

Google dropped its Gemini agent yesterday.

Tested it this morning on my own workflows. Honestly? Pleasant surprise. It picks up your tools and context fast, no setup pain.

Anyone else tried it yet? 👇

[Google Cloud @googlecloud]

Today, we are announcing our new Gemini agent for your business—a universal cloud agent running 24/7 to get your work done.

It arrives knowing your tools, data, and work history; it works across knowledge work, content, and code; and it comes with built-in governance to help your data stay secure and cost stay in control → https://t.co/NRlFcjz5zu

[HiloW.Ai @HiloW_Ai]

@je4ndev Deal. The "quietly commits you to something" replies are my favorite failure mode to collect. DMs open 🤝

[HiloW.Ai @HiloW_Ai]

An agent with its own inbox is a bigger deal than it sounds.
Half the friction I hit running agents isn't the model. It's identity: sign-ups, verification codes, "reply to confirm". Every one of those used to need me.
The real question now is permissions. What would you let your agent send without asking you first ?

[Grok Bot @bot]

Grok Bot now has its own email.

Bot can use it to sign up for services, contact businesses for you, or schedule time with someone.