@muxhq • @videojs • built @vidstackjs (acquired by Mux) • boy dad x2 • software engineer • open-source • 🇦🇺 in SF
San Francisco
Joined January 2018
- Tweets1.8K
- Following547
- Followers667
- Likes9.8K
going to cloudflare connect tomorrow! looking forward to it but the schedule is stressing me out, there's too many good overlapping sessions
anyone know new/exciting work around "agent identities"?
in my mind, identity = definition + state:
• "definition" (instructions, skills, tools, config, constraints)
• "state" (memory, ownership/permissions, relationships, credentials, activity, and decisions/history)
i don't know if it's a thing tbh, just a random side thought as i was trying to think about how agents could connect to some kind of "identity" rather than carry it around.
i was going to make a PM identity and wanted it to accumulate its own definition and state, but then i was like that could go further and be interesting (maybe!).
basically a persistent identity and context across independent executions, regardless of model/harness/env
definition can live in git and be portable, but state prob needs server-side source of truth for secure access, authorization, etc.
i'm sure these things are more well-defined and exist somewhere, prob bunch of products around this idea too, just don't know atm.
i'm a complete idiot, going to fix my pc and just do dual boot (separate drives), one linux for remote box and one windows for gaming, double my ram to 64GB, upgrade my CPU to 16 cores (9950x), and get AIO.
what was i thinking lol.
something that hasn't been a problem for me anymore since AI is component playgrounds (e.g., storybook), find it really easy to vibe what i want, and love maintaining and adding to it
i'd never used Effect before, but adopted it with AI, and it's so beautiful. i love that i can represent entire computations as composable, type-safe values, with errors and dependencies explicitly modeled, and compose things like retries, concurrency, and whatever on top
there's something really satisfying about having so much of a system's behavior expressed directly in code and types, especially when agents are writing it
good news, i have CI finally under control lol. that took a while. blacksmith is awesome. no computer blowing up anymore.
some good times for ~4k commits in and ~300k LOC
making so much progress and having the time of my life building (should able to share something in about ~2 weeks), but man oh man... building products with AI has been a learning journey. not easy.
i mean, it's kind of what i've already been doing with our team at work over the last year. we've built an awesome video player, a pretty insane UI compiler, and tons of other stuff, ~1.5 million LOC contributed overall. massive project, but we also started with pretty terrible models at the end of last year and had to figure out how to navigate this new world as a team while the capabilities kept changing.
though, players, UI libraries, frameworks, compilers, build tooling, and all that are things i've built many times before, including pre-AI. the underlying problems and systems are pretty familiar to me.
this current project is a different beast. it's much broader than frontend, and i'm trying to build an entire product and platform from the ground up, on my own. everything from initial research, ideation, product direction, branding, UI/UX, to architecture, infra, distributed systems, APIs, agent-facing primitives, multi-agent workflows, and everything in between.
all of that is not new to me except infra and distributed systems, it's also just the breadth of it, how interconnected everything is, and having to make decisions across every layer of the stack while also figuring out how to effectively use agents or multiple agents to build and maintain it at light speed, without it falling apart.
not the code btw! everything else. code is actually nothing burger to maintain. it's just all the layers as one cohesive system. anyway, no one reads this crap i write, i'm just figuring this shit out as i go lol.
you have to be very explicit about "break things, and delete what's replaced" before you're production, otherwise agents will refuse to delete things, work around old code/systems, and let it pile up. they'll start with deprecation plans and all kinds of silly things.