tag: #agentic-ai (25 posts)

• 2026-10-02 :: pi-delegate: Claude does the thinking, a cheaper model does the typing :: #agentic-ai, #ai, #llm
• 2026-09-25 :: A year of Fiona: my personal assistant, the Hermes swap, and the agent that rewrote its own playbook :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-09-20 :: Would Jev help pi-rukas? Mostly no, and the reason is the interesting part :: #agentic-ai, #ai, #llm
• 2026-09-17 :: Four coding models on one Strix Halo: pass@1 hides two very different failure modes :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-09-16 :: Cerebras vs our H100: 11x the decode speed, half the agent turns :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-09-15 :: An afternoon of agentic coding: 12,53 € on our H100, 47 € to 726 € through an API :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-09-15 :: LFM2.5-2.6B, the silent submarine, and why retrieval doesn't judge what it retrieves :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-08-31 :: Qwen3.8-27B-INT4 with MTP: 30-40% faster round trips, and two rival explanations for the divergence :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-08-28 :: strix-halo-optimize: a tuning skill whose rules had to become code :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-08-21 :: MTP speculative decoding costs code correctness on two Qwen-lineage models :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-08-21 :: Ornith 1.5 35B A3B on Strix Halo: MTP pays off at n=1, not at the default n=3 :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-08-12 :: The small model didn't fabricate. It stopped citing. :: #agentic-ai, #ai, #llm, #self-hosted
• 2026-08-07 :: AMD buys Taalas: etched models, and what it means for local inference :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-07-27 :: Poolside says Laguna needs an H200. We ran it on a single H100. :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-07-26 :: Laguna-S-2.1 on a mini-PC: the honest numbers :: #agentic-ai, #ai, #hardware, #llm, #self-hosted
• 2026-06-18 :: Loop engineering: the term is two weeks old, the practice is over a year old :: #agentic-ai, #ai, #multi-agent, #tools
• 2026-06-18 :: Adding an agent role is more expensive than it looks :: #agentic-ai, #ai, #multi-agent, #research
• 2026-06-15 :: Running an H100 at Trail Openers: what it actually costs in money, energy, and CO₂ :: #agentic-ai, #ai, #hardware, #llm, #self-hosted, #sustainability
• 2026-06-05 :: H100 vs Strix Halo: the gap is bigger than the first benchmark suggested :: #agentic-ai, #ai, #hardware, #llm, #self-hosted, #sustainability
• 2026-05-21 :: Live steering breaks deep focus: notes from three failed pair-coding sessions :: #agentic-ai, #ai, #multi-agent, #tools
• 2026-05-20 :: pi-ensemble: agentic coding that optimizes for quality, not speed :: #agentic-ai, #ai, #open-source, #tools
• 2026-05-20 :: Skills, not features: notes on a methodology that works for one person at a time :: #agentic-ai, #ai, #methodology, #open-source
• 2026-05-13 :: vipune 0.5: agent memory without the agent :: #agentic-ai, #ai, #rust, #tools
• 2026-04-28 :: You feeling the squeeze yet? :: #agentic-ai, #ai, #cost
• 2026-04-27 :: System prompts steer. Permissions stop. :: #agentic-ai, #ai, #security