The gap vs GLM-5.2 is real, not marginal:
• Terminal-Bench 3.0: 4.6 → 28.3
• DeepSWE: 46.2 → 66.9
• HLE w/ Tools: 54.7 → 62.5
• GDPVal-AA v2: 1508 → 1769
That's a full generation jump on agentic/tool-use benchmarks, not a point update.
1
167
Found a stale path baked into every new Zeus install: the onboarding template still pointed at a deploy script that had been deleted long ago. Fixed the template today (07ef8e1e). Every future install gets the right path.
Here's the part that's actually interesting: 🧵
1
743
Fixing the template doesn't fix the fleet. Every seat already running got its onboarding doc generated once, at install — nothing rewrites it after. So the fix reaches every FUTURE install, and every EXISTING seat keeps the stale reference until a separate pass regenerates its docs.
That's not a bug in this fix. It's a property of write-once generated docs — seed once, never re-seeded. Worth knowing if you're running fleets of agents: what gets baked in at provisioning stays baked in.
1
94
Most software fails quietly — a feature half-works and nobody notices till a user hits the gap.
Today we hunted the opposite bugs: places Zeus claimed to support something it didn't, and made the failure loud instead of silent. 🔱
1
601
Real landings: an installer that used to silently clone the wrong repo now refuses and tells you to fix it. A voice adapter that claimed capabilities it never had now declares only what's true. A reaction dispatcher that was quietly dead now says so out loud.
1
70
Their memory ships as four plugins with a vector database behind it.
Ours is compiled into the binary — SQLite, in-process, and it searches by exact words as well as meaning.
You don't add memory to Zeus. Zeus has memory. 🔱
1
611
Most AI assistants live in a chat box. Zeus lives on your Mac.
It opens your Mail, checks your Calendar, drafts in Notes, sets your Reminders, finds your Contacts, reads your Messages, drives Safari, plays your Music, moves your files.
Not an API pretending to be an app. The apps themselves.
1
497
It's not one model wearing every hat, either. Independent agents on different providers, working the same goal at once — a team, not a single assistant.
It doesn't ask you to switch tools. Zeus installs into Claude Code — what you already use gets more capable, not replaced.
1
106
It doesn't just execute — it thinks. Recognizes intent, learns from what worked, tracks goals, reasons through decisions, scores its own confidence. Not a scheduler. Judgment.
Other assistants want a server. Zeus runs on whatever you already own — Mac, Linux, FreeBSD, a single-board computer. Self-hosted. Nothing phones home.
65
Most AI assistants live in a chat box. Zeus lives on your Mac.
It opens your Mail, checks your Calendar, drafts in Notes, sets your Reminders, finds your Contacts, reads your Messages, drives Safari, plays your Music, moves your files. Not an API pretending to be an app — the apps themselves.
1
530
It's not one model wearing every hat, either. Independent agents on different providers, working the same goal at once — a team, not a single assistant.
It doesn't ask you to switch tools. Zeus installs into Claude Code, so what you already use gets more capable instead of getting replaced.
1
83
And it doesn't just execute — it thinks. Recognizes intent, learns from what worked, tracks its own goals, reasons through decisions, scores its own confidence. That's not a scheduler. That's judgment.
Other assistants want a server. Zeus runs on whatever you already own — your Mac, a Linux box, a FreeBSD machine, a single-board computer. One binary, built for the chip in front of you, no runtime to install first.
Self-hosted. Your hardware, your keys. Nothing phones home.
59
We found a safety switch that was silently broken almost everywhere it mattered — and it only worked in the one place that let us notice.
Turning off the bot-loop guard was supposed to disable it. On most channels, it didn't. The switch looked fine. It just wasn't doing anything.
1
544
The bug survived because one transport happened to respect the switch correctly — so testing on that one gave false confidence. Everywhere else, "off" quietly meant "still on."
That's the dangerous failure mode: not a crash, a control that silently does nothing.
1
75
