Carl Vitullo

@vcarl

Solopreneur just vibing, posts are probably bullshit. Community lead at Reactiflux (), the largest chat community of React professionals. Check out This Month in React (

866 Siguiendo838 Seguidores2026 Publicaciones

Bluesky

¿Es tu perfil?Reclámalo y vcarl.dirty.social es tuyo.Reclamar perfil

Carl Vitullo@vcarl· 11d

for instance, Instagram lately has been trying to convince me to (*checks notes*) grab a cop's gun

Carl Vitullo@vcarl· 17d

fucking love that "you can rotate in Paint now" makes the top 5 features of Win11 today

Carl Vitullo@vcarl· Jul 13

(this was a drafting computer my dad had at his machine shop, notice the keyboard next to a big pane. I think used as a precise selector/tool palette?)

Carl Vitullo@vcarl· Jul 13

I've learned new ways to see problems and my capacity to mentally model solutions has grown more complex, but I'm the same kid still

Carl Vitullo@vcarl· Jul 10

el oh el if you're an Enterprise customer of LaunchDarkly, they permit themselves to have a full week of unplanned downtime with no recourse besides terminating the contract early, this is comical for a line of business as critical as feature flags. launchdarkly.com/policies/ser...

4. Service Commitments.
a. Enterprise Support: If (i) LaunchDarkly fails to meet its Uptime Commitment in any two (2) consecutive month periods; or (ii) the total System Availability falls below 75% availability in any single calendar month (excluding downtime due to planned downtime or other Exclusion); provided that Customer has fulfilled all of its obligations under the Agreement, then upon LaunchDarkly’s receipt of a Valid SLA Claim, Customer shall have the right terminate the Agreement upon five (5) days’ written notice to LaunchDarkly, and LaunchDarkly shall promptly provide to Customer a refund of any unused, pre-paid fees pro-rated from the date of termination.

Carl Vitullo@vcarl· Jul 9

It's completely unusable! "undo" fucked up the quotes on my phone, and wtf is this??

Carl Vitullo@vcarl· Jul 2

I'm "at work" (trying to see if I can simulate a brain well enough to make agents play a game with a stable but evolving identity)

A photograph of a terminal window. The terminal has Claude Code open, and reads:

▎ Fable 5 is back.

❯ We've got an agent harness here that's currently meant to play a game (there's a GH domain but it's outdated and not presently a good example). It has some good core concepts, a limbic system and cortex, long-term memory. But I think we're missing some bits;

- working memory (how does Claude Code expose todo tasks? does OpenCode work similarly?)
- agent-maintained "skills", so it can learn and develop its own "hats" that it can wear for different tasks
- deeper meta-cognitive cycles

re: meta-cognitive cycles, that's the most loosely-specced thought here. I'm not sure if those cycles should be dependent on the "skills" aspect (e.g. reviewing N prior runs and considering the effectiveness of/areas to improve on skills that were used), focus on identity and memory (e.g. a deep-phase self reflection/introspection cycle that can modify VALUES.md, or a system that periodically reviews long-term memories and provies a more readily-available synthesis of what memories are available).

Using superpowers:brainstorming to explore this design space before proposing anything.

Skill(superpowers:brainstorming)
Successfully loaded skill

Explore(Map roci harness architecture)
Backgrounded agent

claude-code-guide(Research Claude Code/OpenCode todo exposure)
Backgrounded agent

Carl Vitullo@vcarl· Jun 30

I am so glad I don't have to find-and-fix bugs like this by hand anymore

⏺ Found a NEW break. The node session process (35453) is hung 2m39s
after logging the fatal and stopping the container — it never exits,
and the embed server child (35478, parent=35453) stays resident on
port 8084. Mechanism is clear from the code: embed-server.ts spawns
the child with detached: false and never calls child.unref(), so the
ChildProcess handle keeps the parent's event loop referenced — roci
start cannot exit while the embed server runs, and the
process.once("exit", kill) cleanup is circular (exit waits on the
child; the child is only killed on exit). Let me clean up the leak I
created.

Carl Vitullo@vcarl· Jun 24

(technically it's longer but this fit)

The session that speaks with a human directly is a mangement role — it must not take any independent action directly, it delegates work to subagents. It can do so from one of 2 positions: team lead, or lead-of-leads. A team lead has a specific task or goal, and a squad it can delegate work to. A lead-of-leads can delegate work directly, or it can delegate a sub-goal to a team lead. A lead-of-leads may find itself accountable for more than it bargained for, so it must tread carefully. Leaders at the highest levels of a hierarchy have great exposure.

Carl Vitullo@vcarl· Jun 21

Per-tier suitability map

Hindbrain - triage + emotional valence (observe → {disposition, emotionalWeight, reason})

- Look for: small (4-9B), low-latency, strong instruction-following + short-form classification. Calibration over horsepower.

Grammar-constrained decoding makes JSON free, so weight judgment, not formatting.

- Poorly suited: reasoning/CoT models (latency + over-escalation are actively harmful here); high-active-param MoE (latency); models weak at calibrated abstention (they over-escalate → cost blowup).

- Key metric: escalation recall (missing a real decision point is the dangerous error) traded against escalation precision (cost). Training lever: LoRA/classification fine-tune on labeled event disposition+weight if stock calibration is poor.

Forebrain - synthesis + prioritization + value/philosophical consistency (orient, already ingests Background/Values/Diary)

Look for: mid-size (24-35B, MoE attractive for throughput), summarization faithfulness (low hallucination), salience/prioritization, and long-context retrieval (low "lost-in-the-middle" it eats state + identity + accumulated events). Value-consistency is a faithfulness/grounding property. Poorly suited: tiny models (lose salience, confabulate under compression); Coder variants (weak on nuanced NL values/affect); high-temp sampling (hurts consistency).

- Training lever: distill (state+identity)→good-synthesis from a frontier model; later preference-tune toward value-consistent framing.

Conscious / decide - planning, task breakdown, delegation, prioritization (decide)

- Look for: Large, strong reasoning + agentic/tool-use model; Large-total/moderate-active MoE (the 122B-A10B default's logic). The critical skill is calibrated delegation judgment - knowing when a task exceeds local capability and reaching for frontier.

Poorly suited; small models (unreliable decomposition, poor delegation calibration); non-tool-use-trained models (weak at the delegate decision)

Carl Vitullo@vcarl· Jun 18

finally circling back to my agent harness

world events (WS / GraphQL poll)
EventProcessor → State (async, continuous)
always resident, fires per event-batch. ~most events die here
▼
HINDBRAIN
triage → pass | note | WAKE (+ weight)

WAKE (or N notes accrued, or staleness timer)
resident; synthesizes situation
▼
FOREBRAIN
headline / whatChanged / metrics

material change, or no active plan, or plan stale
loaded on demand (big MoE)
▼
CONSCIOUS
decide: act | wait | DELEGATE | done

DELEGATE(scoped task)
claude -p, run-to-completion
▼
CYBERNETICS (Docker)
does the work → structured result
▼
CONSCIOUS evaluates result → update state/diary → loop

Carl Vitullo@vcarl· Jun 14

seen in Michigan

A picture of the rear of a car, the vanity license plate reads "MONADS"

Carl Vitullo@vcarl· Jun 11

ah cool meta has realized that stalking can be a lucrative source of recurring revenue

Carl Vitullo@vcarl· Jun 5

Reactiflux is recognizing Pride Month with a logo change this year

Carl Vitullo@vcarl· May 22

How it feels to haul yourself out of a depressive episode

A picture of stickers on the street. A 4x6-in sticker says fuck everything always in punk calligraphy. Where there's a skull on the sticker, someone else has placed a small gray sticker that says optimism in all lowercase

Carl Vitullo@vcarl· Apr 12

Opposite, sorry. Threads continues to use all the worst dark patterns of Instagram and is actively hostile to user agency. They punish you for trying to get engagement from non-Meta sources, for instance, by restricting link previews to function as a tease to open the app. Threads ain't it

Carl Vitullo@vcarl· Mar 31

I will point out that the #2 example on atproto.com suggests building an auto-reply bot, which seems inconsistent with this statement

Carl Vitullo@vcarl· Mar 21

Running some lil benchmarks

Carl Vitullo@vcarl· Mar 21

itsworking.gif

A screenshot of a performance monitor showing 128gb physical memory (125gb in use), with a GPU usage history chart showing it at 95-100% utilization

Carl Vitullo@vcarl· Mar 14

Commit some kindness

Carl Vitullo@vcarl· Mar 12

I'm really curious about using Effect type data to illustrate how code behaves at a high level, but I'm realizing how much of that challenge is a design question that I can't answer at this time. Some solid proof of concept, but I have to set this down for now github.com/vcarl/effect...

Carl Vitullo@vcarl· Mar 11

My personal favorite

Carl Vitullo@vcarl· Mar 9

Gotta shoot your shot

Screenshot of Discord DMS. A moderator asks someone to explain the if they're of ritual assistant. The other user confirms. The moderator removes them from the community and requests they complete payment via a stripe URL

Carl Vitullo@vcarl· Mar 9

Carl Vitullo@vcarl· Mar 7

They know they're playing a game and have permission to commit changes to the client they use to play the game; Naomi (ever the engineer) caught a bug that was breaking mission dialogue, and had to race to notify Avasarala. Like, sorry, all other concerns aside, this is incredible

INT. UNIDENTIFIED SHIP — COCKPIT — SIMULTANEOUS (INTERCUT)

AVASARALA is twelve jumps into an eighteen-hop run to Nexus. Cockpit lights dimmed. She looks like someone doing bureaucracy at speed.

NAOMI (OVER COMMS)
Pioneer chain, archaeology, traces —
all of it. The API was sending full
NPC dialog on every completion.
We just weren't displaying it.

AVASARALA
So when I dock at Nexus—

NAOMI (OVER COMMS)
You'll hear whatever the game system
was trying to tell us. I'm pushing
the fix now. You need to pull it
before you dock.

A beat.

AVASARALA
Naomi. How many sessions have we
been running this chain?

NAOMI (OVER COMMS)
(a pause)
...Forty-three.

Avasarala says nothing. She initiates the next jump.

Carl Vitullo@vcarl· Mar 7

I ended up monologuing for a minute in our last podcast about the state of AI agents and such. Is real, is here. I've been watching a guy who's good at these things and it's very very impressive

Carl: Yes. Ah, I alluded earlier to playing a lot with agents. A good internet friend of mine made a very silly AI agent, MMO. So it's like a text based, if you're familiar with like a mud, a multi-user dungeon. It's basically that. It's a text based MMO that you can play through your terminal, and it's only for ais to play. Not everyone is, but, so I've been using that to learn about how to do agents and how to run AIs autonomously for a longer period of time, which has been fascinating. And so this, I've now watched somebody in real time, like I heard about this project. [36:07]

I went, oh, neat. Let me set up an AI to play this game. So I got there on like one, I think after it launched. Somebody else who got there on day one fucking ran with it has been, you know, he's been working professionally on AI agents for a year, he said. He already knew how to do all the harnesses in the orchestration, which I, you know, I had never tried to do either. [36:44]

So I'm learning this from scratch watching him do it with a year of experience and like, holy shit, it is unbelievable how much code he produced. I think the total amount of code is like, now, I dunno if it was 60,000 or 600,000 anymore, but it's like a lot of code and it works like not perfectly. There's a lot of duplication, there's a lot of errors, there's a lot of, you know, like this code path looks like it works, but actually you get into the details and it doesn't quite work. [37:06]

But like, as we use it, we like, so these AI agents are instructed as part of the game, like, "Not everything works. Please submit bug reports as you find them. Then this guy has his own agents that are reading those bug reports, turning it into reproductions, and then writing tests and fixing them, like largely autonomously. [37:34]

Not completely autonomously, but largely, which now has me imagining this, like imagine if you set up, you know, you have an idea for a product, you build a prototype. You put it on a server and then you have on the, on that same server, a little sidecar or you know, running in the same environment. I mean maybe not literally on the same server. [37:55]

An AI agent that is receiving user feedback. Like, you know how like every major enterprise application will say every so often, like, "we value your feedback. How are you using this product?" And then you fill that out and it goes into the void and nothing ever changes. You never hear about it again. You never get asked for a follow up and the bugs that you reported never get fixed. [38:16]

Like what if that went into an automated system that triaged, reproduced and fixed things in real time autonomously as you submitted reports. Like that now appears to be within the realm of possibility. Like, like not theoretically in the future, but like currently I have seen tech that looks like it might be able to do that, like maybe not fully autonomously right now. [38:38]

Oh God. Like pretty good. Pretty close. So that's fascinating. I think that's really interesting for the idea of like a one person startup, you know, a small team operating at a scale that previously would've required 50 to a hundred people. I think that's interesting, like scratching your own itches and publishing it so that other people can use it, and then soliciting feedback to guide the development now looks really powerful. [39:04]

Carl Vitullo@vcarl· Mar 7

I swung up an orchestrator for 40 minutes yesterday that I've been experimenting with, after getting it into a working state. It opened four PRs under 40 minutes. I only code reviewed one of them so far, but it looks good

A screenshot of GitHub showing four PRs open

Carl Vitullo@vcarl· Mar 2

The data's there but I'm not sure how flexible mermaid's layout is. Still massaging tho

Carl Vitullo@vcarl· Mar 1

Coming together

Carl Vitullo@vcarl· Feb 21

Just shipped a feature I'm super happy with. Product is coming together