Was just reminded of mIRC scripting by @ianlivingstone. Man, that's a blast from the past. I was INTO that.
Variable variables in the programming language as mock arrays was _something_.
People often talk about constraints begetting creativity. Classic example!
This sort of feedback is priceless.
I wonder if this is a sort of inevitability of RSI; with AI as the challenger and judge in RL, then perhaps models asymptote towards 'agentese'.
More capable, but increasingly dissociated with the human experience.
Theo - t3.gg @theo 19h Why I hate GPT-6.2 Astra During early testing, I found myself so frustrated with this model that I spent a day writing a letter to OpenAI about it. They fixed a small handful of my complaints, but not enough for me to
I love that this is now a polished native feature of @OpenAIDevs Codex. You can annotate any of the agent's previous comments and have a very precise way of giving feedback. I use it SO MUCH.
Guess I was onto something ... 😜
This extension is a game-changer for me.
1. Agent sends long message, I `/feedback`.
2. I review and annotate comfortably in @code (my $EDITOR).
3. I can add additional text around the [Feedback] placeholder in my next msg.
4. Model sees my annotations.
5. Get my sanity back.
OWASP's new agent security guidance reflects a growing consensus: IAM is the fundamental control loop for agents, enforced outside the model itself.
Agents should be free to reason and iterate, but their authority must be independently governed.
genai.owasp.org/resource/sec…
the entire advantage of llms is they are incredibly flexible
which means they can conform to whatever way you prefer to work and think and build
there's infinite people posting about the best way to use agents. like, first you need to do a plan, and then you need to blah blah blah
but if that doesn't click for you, don't work that way. work however you like
the little bit of advice i'd give is think about ways to keep it fun
it can get boring consuming big chunks of text, ask for a fun video in a crazy voice that gets unhinged over time
I've been trying an experiment.
1. I have a high-entropy vibe pile that mostly does what I want but accumulated so much accidental complexity.
2. I tried to get codex (astra high) to act as a PM orchestrating a bunch of 'clean room' agents to first design specs (code and prose) and then implement against them.
The delta in total code is huge (way less lines in the rewrite) but in a system of this complexity the agent-isms creep in so quickly.
"Well what if this database write fails after a timeout during rare-operation-xyz?" 10 new test suites, a whole new set of bookkeeping and complexity, repeated fractally everywhere.
Still haven't found the magic bullet to keep entropy under control.
Just like the market expectations rising due to agent assistance, my expectations of agents (and models) keeps going up.
Amazing how quickly I can find even the latest generations lacking when they would have been inconceivable just a few months ago.
Durable agents need auth that survives the pause. @KeycardAI built an interceptor for the Temporal Python SDK that fetches a fresh, scoped token before every activity attempt and keeps tokens out of workflow history.
Write-up and @aiDotEngineer NYC demo: keycard.ai/blog/keycard-temp…
Here's the recording from AGNTCon.
It's advertised as a debate but I feel like we all forgot what point of view we were supposed to be supporting 😅
First time on stage in years!
youtu.be/gEiM4HBtPAU?si=07Hw…
Surprise! Turns out I'm subbing in for our fearless @KeycardAI leader, @ianlivingstone for a panel this afternoon at AGNTCon: "Security in a Headless World".
Will be both exciting and intimidating to be on the stage, debating with giants like @dexhorthy, @scottcjohnston and @andrewbecherer.
Wish me luck! 😅
.@ajambrosino agent-> human prompts can easily exceed the vertical space of the ChatGPT window (and screen) making it impossible to scroll to see the full question and the context in which the question is being asked.
Suggestion: pop-out option or open-questions sidebar / ui.
And not just reason but also have facilities for safely changing those facts through ITL other harness-facilitated flows.
Bringing something like this to MCP would be a huge unlock for mission-based authorization.
Credentials and corresponding challenges need to be in-path of a harness so that an agent can reason over its own access paths. Challenges in particular are useful here.