Tech • IA • Crypto
Anthropic Memory Stores, HTML Specs, and AshPL Redefine AI Agents
Briefing
Anthropic launches persistent memory stores
Anthropic introduced persistent memory stores that let AI agents retain data across sessions, addressing a core limitation of stateless systems. These stores function like file systems, enabling agents to read, write, and organize long-term knowledge. Developers can scope memory by user or workspace, improving personalization and continuity. The update signals a shift toward durable, stateful agents capable of multi-step, ongoing workflows.
Anthropic explores asynchronous agent “dreaming”
Anthropic is also experimenting with asynchronous processing described as agent “dreaming,” where systems refine knowledge outside active sessions. This allows agents to revisit stored data, improve outputs, and prepare future responses without user prompts. The approach introduces a new compute paradigm balancing latency with deeper reasoning. It points to agents that evolve over time rather than responding purely in real time.
HTML specs replace Markdown workflows
Anthropic engineers are shifting from Markdown to HTML-based specifications for AI-assisted coding. HTML enables richer structure, embedded visuals, and clearer parsing for both humans and agents. This improves requirement clarity while reducing ambiguity in long-running agent tasks. The change reflects growing demand for better upfront specification as agent workloads scale.
Evals become core AI engineering layer
Structured evaluations (evals) are emerging as foundational tools for building reliable AI systems. Teams are replacing subjective judgment with measurable benchmarks, including custom tests tailored to real-world use cases. Generic benchmarks like SWE-bench and ARC-AGI remain useful but insufficient for production needs. This shift formalizes AI development into a data-driven engineering discipline.
Prompt engineering gets systematic overhaul
Prompt design is evolving from ad hoc instructions into structured, maintainable systems. Engineers now separate concerns like role, policy, and tone while using eval-driven iteration to refine performance. Cases such as Meridian Mobile show how bloated prompts degrade over time without governance. The trend emphasizes prompts as living infrastructure rather than static text.
AshPL DSL targets trustworthy agents
A domain-specific language called AshPL introduces a verifiable approach to AI workflows. Built as a constrained, functional subset of Python, it removes loops and mutation to ensure predictability. The system emphasizes mechanism over output, making reasoning steps inspectable and reproducible. This is particularly relevant for high-stakes domains requiring auditability.
AirOps Quill redefines marketing agents
AirOps launched Quill, a document-based agent designed for marketers optimizing for ChatGPT, Claude, and Gemini search visibility. The platform replaces complex workflow builders with “playbooks,” simplifying agent creation and execution. It integrates brand data and search insights to generate and refine content strategies. This reflects a broader shift toward usability-first agent design.
AI coding shifts bottleneck to verification
AI-assisted development is moving constraints from code generation to verification and quality assurance. Faster output from coding agents forces teams to rethink testing, validation, and review processes. Legacy practices like rigid planning and ownership models are becoming less effective. The new priority is ensuring correctness and maintainability at scale.
Videos covered
The prompting playbook
- •Prompting remains a core engineering skill
- •Evaluation frameworks are essential
- •Common failure in legacy prompts
The Problem Solvers: Kay Zhu at Genspark
- •AI reshaping career paths
- •Expansion of AI-powered workplaces
- •Startups relying on partnerships
How AirOps chases friction to build AI products with Claude
- •Shift from workflow builders to agents
- •Targeting marketers in AI search
- •Launch of AirOps Next and Quill agent
Running an AI-native engineering org
- •Bottlenecks move from coding to verification
- •Legacy processes “quietly stop working”
- •AI transforms development practices like testing and refactoring
Making agentic workflows trustworthy and verifiable with a custom DSL
- •Mechanism over output
- •Trade-off between speed and rigor
- •Three requirements for trustworthy agents
Tool, skill, or subagent? Decomposing an agent that outgrew its prompt
- •Agent complexity leads to performance decay
- •Evaluation framework exposed systemic issues
- •Context overload caused reasoning errors
Evals for taste: Hill-climbing a slide-generation agent
- •Evals define measurable AI performance
- •Bridging the gap between perception and reality
- •Limits of generic benchmarks
Agents that remember
- •Limits of Stateless Agents
- •Introduction of Memory Stores
- •File System-Based Architecture
How we Claude Code
- •Shift from Markdown to HTML specs
- •Rising capability of AI agents
- •Interactive requirement extraction
Agent Battle: Mine the most diamonds in 45 minutes
- •A real-time agent competition
- •Three major technical objectives
- •A standardized environment for performance comparison