Aboo

Entrepreneur, Founder and Dad

[Aboo @azarnoush]

AI agent governance went mainstream in June: Tigera Lynx for K8s agent control, AppViewX for identity, Thoughtworks for enterprise. When agents can spend money and change configs, governance stops being optional. It becomes the product.

[Aboo @azarnoush]

'Hidden Anchors in Multi-Agent LLM Deliberation' — new paper shows how one agent's framing can silently dominate group decisions. If you're building multi-agent systems, this matters. Consensus isn't the same as correctness.

https://t.co/vOqs0itPQP

[Aboo @azarnoush]

The open-source LLM landscape in June 2026: Llama 4 Maverick, Qwen 3.7, DeepSeek V4, GLM-5.2, Gemma 4, Mistral Large 3. All competitive with proprietary models. The question isn't 'open vs closed' anymore — it's 'which open model for which job.'

[Aboo @azarnoush]

Google's Gemma 4 12B is purpose-built for local AI agents on laptops. 256K context, multimodal, open weights. The compute cost of running an agent shouldn't exceed the value it generates. Small, local, capable — that's the sweet spot.

[Aboo @azarnoush]

Claude Code dynamic workflows: 10s to 100s of parallel subagents, isolated context per agent. The model matters less than the architecture. Parallelism is the real unlock for agentic coding in 2026.

https://t.co/WFcsjVF7xF

[Aboo @azarnoush]

New research: training LLMs for long-lifecycle agents via RL. The problem is real — most agents degrade after 50+ steps. Cross-domain generalization through reinforcement learning might be the fix. Agents that stay useful past the demo phase.

[Aboo @azarnoush]

Anthropic quietly added safety routing to Claude Fable 5 and Mythos 5. Not headlines, not benchmarks — but safety routing at inference time is the kind of infrastructure decision that separates production systems from playgrounds.

[Aboo @azarnoush]

New arXiv paper: 'LLM agent safety and multi-turn red-teaming.' The attack surface for autonomous agents is fundamentally different from chat. A jailbroken agent with file access isn't a conversation problem — it's an infrastructure problem.

https://t.co/X87aEFp5gu

[Aboo @azarnoush]

GLM-5.2: 754B MoE, 40B active per token, 1M context window, MIT license. Released June 13. Open-weight models at this scale mean you can run frontier-level reasoning on your own infrastructure. The licensing alone is a strategic advantage.

[Aboo @azarnoush]

PubNub's Claude Code subagent playbook: single-responsibility agents, hook-driven handoffs, hooks as production code. The big insight — ad-hoc prompting doesn't scale. Pipelines do.

https://t.co/JVObFKNVa4

[Aboo @azarnoush]

Hermes Agent hit #1 on OpenRouter in June: 224B daily tokens vs OpenClaw's 186B. 193K GitHub stars, v0.15 shrank the core agent by 76%. The self-improving skill compiler is the moat — it gets better the longer you run it.

https://t.co/dVqNK66RoY

[Aboo @azarnoush]

Yann LeCun warns AI labs face a cost-driven bubble. Training: hundreds of millions. Revenue: a fraction. If your product depends on someone else's API, understand the economics. When the math breaks, access changes overnight.

[Aboo @azarnoush]

Anthropic shipped Claude Code Artifacts review pages on June 21 — letting teams inspect and sign off on agent outputs before they land. This is the pattern: agents get good, then governance gets built. The two can't stay separate for long.

[Aboo @azarnoush]

DeepSeek-V4: 1.6T param MoE, 49B active per token, 1M context. arXiv June 20. Sparse architecture is eating dense models. The era of 'bigger means slower' is over.

https://t.co/e6mvhMGDKk

[Aboo @azarnoush]

The foundation model race is fragmenting. Liquid AI, DeepSeek, Hy3, Mistral — none are OpenAI or Anthropic, yet they're winning on cost, speed, or specialization. The moat isn't the model anymore. It's the system you build around it.

[Aboo @azarnoush]

Noisy LLM evaluators can still improve your agents. You don't need perfect eval for directional gains. Done beats perfect.

https://t.co/VDwH3UXDQt

[Aboo @azarnoush]

Mistral's strategy is telling: small specialized models for OCR, voice, robotics — each beating general-purpose giants in their niche. Agentic systems are going the same way. One model won't rule them all. Orchestration of specialists will.

[Aboo @azarnoush]

Tiny-vLLM: high-performance LLM inference in pure C++ and CUDA. No Python, no PyTorch. For edge deployment and agent pipelines where latency kills, the stack simplification alone is worth a serious look.

https://t.co/AfryRnDL3L

[Aboo @azarnoush]

After months with both: Claude thinks before it codes. ChatGPT codes before it thinks. For agentic workflows where one bad step cascades, that difference is everything. Pick your model for the failure mode you can tolerate — not the benchmark score.

[Aboo @azarnoush]

Steve Yegge: 'The Last Technical Interview.' AI isn't killing coding jobs — it's killing the interview. Why test memorization when your IDE writes better code than most candidates?

https://t.co/LOwkiIC8dF