@NousResearch Thanks! Planning an MCP server in a sibling container that holds the limited Gmail creds.
Untrusted input is the hard part I think. Usually I would extract structured intent per mail, but I want Gmail as a knowledge base for Hermes, so it needs full text. How do you handle that?
Gmail + Hermes Agent: how do you limit prompt injection risk? Looking for read + draft-only access, no send/delete. What's your setup? @NousResearch#HermesAgent#AIagents
2 days digging through CI logs and transitive dependencies. Result: a 1-line fix.
That's how this story would have gone.
Instead, AI found it in 60 seconds. I skipped the AHA moment, and honestly? Happy to leave that part of the job behind.
#SoftwareEngineering#AI#CICD
Pizza's bestellen? Dat had de AI ook gekund. 🍕
Het bewijs: een AI-agent bestelt zelf de pizza's via MCP + Playwright.
Met UCP en AP2 voor het afrekenen kan het straks volledig automatisch.
Over de training: https://t.co/nFozeycHxt
#AIAgents#MCP#Playwright#AgenticCommerce
@joelhooks@ryanvogel Random software NPC here, but I think "vibe coding" is still a pretty clear term: nobody seriously looked at the code, only at the output (which can be fine for your use case).
Agentic engineering is the opposite: you babysit your harness and review what it builds.
I feel like we've been anticipating a cataclysmic cyber event for so long that we've worked ourselves up to responding too heavily, with site takedowns rather than prudent SecOps until #PHP2Shell patch day.
So much work tuning context, skills and MCP, for things the bigger models increasingly do out of the box.
Let's hope it pays off, and the good models don't get cheap enough to make all these new skills worthless. 😅
#ContextEngineering#AgenticAI
@dexhorthy Better planning and prompting let smaller models do the job. When the VC subsidies run out, we'll be glad our toolchain can run on small open-source models.
@csswizardry It's just that we use it too much. Claude was a breath of fresh air after ChatGPT's annoying personality, but now I'm sick of "let me be real with you", "honest", and "here's the thing". As consumers, custom primary agents with a personality brush-up can help.
How do you all fit #Qwen into your development workflow? I mostly work with Luna. Qwen3-Coder-30B could handle some of the minor tasks, but I tend not to switch because of the context-switching cost. Do you use a smart router for that? #LocalLLM
Continuing my wayfinder skill adventure: the prototyping skill it uses is awesome! Multiple #mockups presented neatly side by side. Thanks @mattpocockuk !
Enjoying #grillme? Try @mattpocockuk's wayfinder skill: it runs deep research in parallel between grilling sessions, while you stay in control.
The exploration counterpart to Sandcastle's #softwarefactory, with the control I found hard to get from basic setups.
#coronavirus spread over time visualized using interactive map: https://t.co/fxyLgyKDtv #gis. Infection is less widespread than other maps suggest. Map design blog post https://t.co/N9SROFBC4h .
@netflix Please allow us to disable the auto play tiles. It is really annoying having to use an ad blocker to disable it or install and Chrome extension (and switch to Chrome) just to disable spoilers and this intrusive UI.