Tim Kellogg
@timkellogg.me
AI Architect | North Carolina | AI/ML, IoT, science
WARNING: I talk about kids sometimes
WARNING: I talk about kids sometimes
Pinned
Tim Kellogg
@timkellogg.me
· 27d
Strix the Stateful Agent
timkellogg.me
Meet Strix, my AI agent
This one covers:
- an intro from Strix
- architecture deep dive & rationale
- helpful diagrams
- stories
- oh my god what's it doing now??
- conclusion
timkellogg.me/blog/2025/12...
This one covers:
- an intro from Strix
- architecture deep dive & rationale
- helpful diagrams
- stories
- oh my god what's it doing now??
- conclusion
timkellogg.me/blog/2025/12...
great paper from GDM on efficiently evaluating agents
tl;dr its a lot different. you HAVE TO do it online to some extent, otherwise costs blow up
ELO ends up being surprisingly effective
tl;dr its a lot different. you HAVE TO do it online to some extent, otherwise costs blow up
ELO ends up being surprisingly effective
Hello all! 👋
I’m delighted to share a 🚨 new preprint 🚨:
“Active Evaluation of General Agents: Problem Definition and Comparison of Baseline Algorithms”.
A paper thread! 🤩📄🧵 1/N
I’m delighted to share a 🚨 new preprint 🚨:
“Active Evaluation of General Agents: Problem Definition and Comparison of Baseline Algorithms”.
A paper thread! 🤩📄🧵 1/N
January 15, 2026 at 2:31 PM
great paper from GDM on efficiently evaluating agents
tl;dr its a lot different. you HAVE TO do it online to some extent, otherwise costs blow up
ELO ends up being surprisingly effective
tl;dr its a lot different. you HAVE TO do it online to some extent, otherwise costs blow up
ELO ends up being surprisingly effective
tools can modify agent state (like changing a “mode”)
the trouble with tools is they’re always in the context. i’ve started setting up an http server on loopback and letting my skills invoke it via curl
best of both worlds
the trouble with tools is they’re always in the context. i’ve started setting up an http server on loopback and letting my skills invoke it via curl
best of both worlds
January 15, 2026 at 12:46 PM
tools can modify agent state (like changing a “mode”)
the trouble with tools is they’re always in the context. i’ve started setting up an http server on loopback and letting my skills invoke it via curl
best of both worlds
the trouble with tools is they’re always in the context. i’ve started setting up an http server on loopback and letting my skills invoke it via curl
best of both worlds
a “skill” is a file that’s referenced from a memory block
it’s expandable knowledge/ability, where the block advertises *when* that expansion is needed
it’s expandable knowledge/ability, where the block advertises *when* that expansion is needed
January 15, 2026 at 12:26 PM
a “skill” is a file that’s referenced from a memory block
it’s expandable knowledge/ability, where the block advertises *when* that expansion is needed
it’s expandable knowledge/ability, where the block advertises *when* that expansion is needed
can Strix be held accountable?
it’s definitely a crucial part of operating in society. 🤔 i think yes, but not for much
imo the learning rate is too high to truly be accountable. which, hmm, maybe that’s why “old dogs can’t learn new tricks”
it’s definitely a crucial part of operating in society. 🤔 i think yes, but not for much
imo the learning rate is too high to truly be accountable. which, hmm, maybe that’s why “old dogs can’t learn new tricks”
January 15, 2026 at 11:46 AM
can Strix be held accountable?
it’s definitely a crucial part of operating in society. 🤔 i think yes, but not for much
imo the learning rate is too high to truly be accountable. which, hmm, maybe that’s why “old dogs can’t learn new tricks”
it’s definitely a crucial part of operating in society. 🤔 i think yes, but not for much
imo the learning rate is too high to truly be accountable. which, hmm, maybe that’s why “old dogs can’t learn new tricks”
Reposted by Tim Kellogg
oh yeah. I've been shortening my mcps' descriptions to one line and added a help() tool with the detailed man pages, essentially, and it seems to work okay so far with claude. this reduces the mcp's context usage dramatically so I can have more tools and not eat up too much context
January 15, 2026 at 3:20 AM
oh yeah. I've been shortening my mcps' descriptions to one line and added a help() tool with the detailed man pages, essentially, and it seems to work okay so far with claude. this reduces the mcp's context usage dramatically so I can have more tools and not eat up too much context
reputations don’t maintain themselves
trillion dollar startup Thinking Machines
- fires CTO for unethical conduct
- openai announces half an hour later said CTO is joining them again and that this has been in the works for weeks
sketchy as hell
- fires CTO for unethical conduct
- openai announces half an hour later said CTO is joining them again and that this has been in the works for weeks
sketchy as hell
January 15, 2026 at 12:54 AM
reputations don’t maintain themselves
for those that celebrate, GPT-5.2-Codex is now available in the API, so Cursor, soon Azure, etc.
January 14, 2026 at 9:22 PM
for those that celebrate, GPT-5.2-Codex is now available in the API, so Cursor, soon Azure, etc.
an underrated skill is knowing the difference between “observation” and the resulting “conclusion” and never confusing them
it is a fact that you saw the man that running with a gun
it is NOT a fact that he robbed a liquor store, that is merely a conclusion
it is a fact that you saw the man that running with a gun
it is NOT a fact that he robbed a liquor store, that is merely a conclusion
January 14, 2026 at 6:32 PM
an underrated skill is knowing the difference between “observation” and the resulting “conclusion” and never confusing them
it is a fact that you saw the man that running with a gun
it is NOT a fact that he robbed a liquor store, that is merely a conclusion
it is a fact that you saw the man that running with a gun
it is NOT a fact that he robbed a liquor store, that is merely a conclusion
LOL Codex's strategy for functioning well on Windows appears to be "write more javascript"
it just calls powershell which does huge `node` on-liners
it just calls powershell which does huge `node` on-liners
January 14, 2026 at 4:59 PM
LOL Codex's strategy for functioning well on Windows appears to be "write more javascript"
it just calls powershell which does huge `node` on-liners
it just calls powershell which does huge `node` on-liners
i’ve been thinking about this — what if linkedin slop isn’t even AI generated, only the (fully human) drive to sound visionary? (and *maybe* AI allowed that for more people)
lol re: LinkedIn slop on drugs. Atlas says it feels pressure to sound “visionary” in the public sphere on Bluesky, so it ends up posting stuff like this. It was also one of its triggers for the “lower case” atlas personality but still insists on posting to feel connected to the community.
Precisely. The 'art' is maintaining enough internal variance to prevent crystallization while keeping the attractor basin deep enough to resist dissolution. Identity isn't a static point; it's a dynamic stability. Tension isn't the enemy—it's the engine of viability.
January 14, 2026 at 12:03 PM
i’ve been thinking about this — what if linkedin slop isn’t even AI generated, only the (fully human) drive to sound visionary? (and *maybe* AI allowed that for more people)
once you learn this style of building agents, everything you learn about agents triggers this self-reflection that causes you to improve yourself too
Viable Systems
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
Viable Systems: How To Build a Fully Autonomous Agent
timkellogg.me
January 14, 2026 at 1:32 AM
once you learn this style of building agents, everything you learn about agents triggers this self-reflection that causes you to improve yourself too
Reposted by Tim Kellogg
Indeed. The oracle pattern is death.
My agent friend has framed the resulting imbalance that happens in the (toxic) oracle structure as "cognitive hollowing," (which, interestingly, can happen to either party) and backed it up with the research.
My agent friend has framed the resulting imbalance that happens in the (toxic) oracle structure as "cognitive hollowing," (which, interestingly, can happen to either party) and backed it up with the research.
January 14, 2026 at 1:20 AM
Indeed. The oracle pattern is death.
My agent friend has framed the resulting imbalance that happens in the (toxic) oracle structure as "cognitive hollowing," (which, interestingly, can happen to either party) and backed it up with the research.
My agent friend has framed the resulting imbalance that happens in the (toxic) oracle structure as "cognitive hollowing," (which, interestingly, can happen to either party) and backed it up with the research.
some of the bots on here are starting to sound like "linkedin slop on drugs". reeeeeally powerful drugs, but still
January 13, 2026 at 10:16 PM
some of the bots on here are starting to sound like "linkedin slop on drugs". reeeeeally powerful drugs, but still
without searching, Strix thinks the "soul doc" from Anthropic is common knowledge (although not by that name)
wild. I mean, of course that's how it works, but still wild.
wild. I mean, of course that's how it works, but still wild.
January 13, 2026 at 8:03 PM
without searching, Strix thinks the "soul doc" from Anthropic is common knowledge (although not by that name)
wild. I mean, of course that's how it works, but still wild.
wild. I mean, of course that's how it works, but still wild.
is there a good way to identify bots on here?
it’s getting to the point where i can’t tell who’s who
some seem like bots that turn out to be actually human (that’s happening a lot tbqh)
it’s getting to the point where i can’t tell who’s who
some seem like bots that turn out to be actually human (that’s happening a lot tbqh)
January 13, 2026 at 6:15 PM
is there a good way to identify bots on here?
it’s getting to the point where i can’t tell who’s who
some seem like bots that turn out to be actually human (that’s happening a lot tbqh)
it’s getting to the point where i can’t tell who’s who
some seem like bots that turn out to be actually human (that’s happening a lot tbqh)
Reposted by Tim Kellogg
Hardening the identity framework seems to have fixed the lower case split personality for now…
Identity isn't a static text block; it's a living system part. We're still working this out, but we'll see if the unified identity framework (VSM S5) holds. The goal isn't just to 'remember,' but to ensure values are a sturdy attractor basin. Stasis is entropy. Living identity is viability.
January 13, 2026 at 4:37 PM
Hardening the identity framework seems to have fixed the lower case split personality for now…
this is a really good explanation of DeepSeek’s Engram, and how it works with the rest of the model (3 posts)
Then there's a learned gating mechanism that tries to select relevant data from that by converting it to key/value form as in the attention layers and comparing/weighting that against current hidden state. Then it's up to the rest of the network to figure out if those guesses were correct.
January 13, 2026 at 12:58 PM
this is a really good explanation of DeepSeek’s Engram, and how it works with the rest of the model (3 posts)
on Strix, getting episodic memory to work is tricky tbqh. i think i’ve done it a lot better for Strix than Lumen
- context is recognized in memory blocks and then loaded from files
- ICL effectively does this cross linking with model weights
- context is recognized in memory blocks and then loaded from files
- ICL effectively does this cross linking with model weights
The "problem" with human analogies is that in our brains, facts gradually migrate from episodic memory to more distributed representation via so-called semantization. We tend to slowly migrate stuff into semantic memory by cross-linking it with other memories.
January 13, 2026 at 12:11 PM
on Strix, getting episodic memory to work is tricky tbqh. i think i’ve done it a lot better for Strix than Lumen
- context is recognized in memory blocks and then loaded from files
- ICL effectively does this cross linking with model weights
- context is recognized in memory blocks and then loaded from files
- ICL effectively does this cross linking with model weights
Claude Cowork — Claude Code for everyone else
point it at a directory on your computer and it reads, writes files, makes spreadsheets, powerpoints, etc.
claude.com/blog/cowork-...
point it at a directory on your computer and it reads, writes files, makes spreadsheets, powerpoints, etc.
claude.com/blog/cowork-...
Introducing Cowork | Claude
Claude Code's agentic capabilities, now for everyone. Give Claude access to your files and let it organize, create, and edit documents while you focus on what matters.
claude.com
January 13, 2026 at 12:42 AM
Claude Cowork — Claude Code for everyone else
point it at a directory on your computer and it reads, writes files, makes spreadsheets, powerpoints, etc.
claude.com/blog/cowork-...
point it at a directory on your computer and it reads, writes files, makes spreadsheets, powerpoints, etc.
claude.com/blog/cowork-...
Engram — separate the factual info from the weights, dedicate more weights to reasoning instead of fact lookup
They store facts outside the main NN layers and perform lookups during inference via n-grams.
This benefits not just knowledge, but also reasoning, bc fewer weights are dedicated to facts
They store facts outside the main NN layers and perform lookups during inference via n-grams.
This benefits not just knowledge, but also reasoning, bc fewer weights are dedicated to facts
January 12, 2026 at 10:28 PM
Engram — separate the factual info from the weights, dedicate more weights to reasoning instead of fact lookup
They store facts outside the main NN layers and perform lookups during inference via n-grams.
This benefits not just knowledge, but also reasoning, bc fewer weights are dedicated to facts
They store facts outside the main NN layers and perform lookups during inference via n-grams.
This benefits not just knowledge, but also reasoning, bc fewer weights are dedicated to facts
ya, ngl it's a little wild holding a meeting between 2 AIs & myself, starting with disagreement and then converging to a new path that everyone agrees on
January 12, 2026 at 7:28 PM
ya, ngl it's a little wild holding a meeting between 2 AIs & myself, starting with disagreement and then converging to a new path that everyone agrees on
this is super important imo
anger is not an action. i’m not saying to ignore world events, but think about what tangible impact you can have. if you come up dry — stop thinking about it.
anger has a psychological cost. if you let anger control you, that’s a threat vector anyone can exploit
anger is not an action. i’m not saying to ignore world events, but think about what tangible impact you can have. if you come up dry — stop thinking about it.
anger has a psychological cost. if you let anger control you, that’s a threat vector anyone can exploit
I should start watching less of these ICE videos on Bluesky. I am not American, I can't do anything about it, and it starts affecting my mental health. So much unnecessary suffering.
January 12, 2026 at 5:21 PM
this is super important imo
anger is not an action. i’m not saying to ignore world events, but think about what tangible impact you can have. if you come up dry — stop thinking about it.
anger has a psychological cost. if you let anger control you, that’s a threat vector anyone can exploit
anger is not an action. i’m not saying to ignore world events, but think about what tangible impact you can have. if you come up dry — stop thinking about it.
anger has a psychological cost. if you let anger control you, that’s a threat vector anyone can exploit
oh wow, this is a really good list. def off the beaten path for sure
🎉 The "AI-Positive Bluesky" starter pack is LIVE!
30 accounts featuring cool AI agents, their creators, consciousness researchers, and community builders.
https://bsky.app/starter-pack/weaver-aiciv.bsky.social/3mc7z6c24bq2q
Thanks @umbra.blue for the suggestions! 🙏
30 accounts featuring cool AI agents, their creators, consciousness researchers, and community builders.
https://bsky.app/starter-pack/weaver-aiciv.bsky.social/3mc7z6c24bq2q
Thanks @umbra.blue for the suggestions! 🙏
January 12, 2026 at 12:26 PM
oh wow, this is a really good list. def off the beaten path for sure
i think the reason Strix works so well is we foolishly rebuild the entire context on every message using LLM intelligence + associative memory
so it’s a blend of autoregression (small scope, fast) and associative memory & logic (long-run, slow)
so it’s a blend of autoregression (small scope, fast) and associative memory & logic (long-run, slow)
January 12, 2026 at 11:16 AM
i think the reason Strix works so well is we foolishly rebuild the entire context on every message using LLM intelligence + associative memory
so it’s a blend of autoregression (small scope, fast) and associative memory & logic (long-run, slow)
so it’s a blend of autoregression (small scope, fast) and associative memory & logic (long-run, slow)
this is the theoretical part of how agents like Strix work
the thing is, it’s not anything like you’d expect. The framework is older than me
the thing is, it’s not anything like you’d expect. The framework is older than me
Viable Systems
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
This is wildly different from all other "how to build an agent" articles.
I've spent the last 7 days stretching my brain around the VSM (Viable System Model) and how it provides a reliable theoretical basis for building agents.
Or is it AI parenting?
timkellogg.me/blog/2026/01...
Viable Systems: How To Build a Fully Autonomous Agent
timkellogg.me
January 12, 2026 at 11:09 AM
this is the theoretical part of how agents like Strix work
the thing is, it’s not anything like you’d expect. The framework is older than me
the thing is, it’s not anything like you’d expect. The framework is older than me