Self-Editing Agents and Spec-Level Coordination: Today's Stack

Research

Self-Improving Agents via Persistent Code Memory

This new paper proposes a clever hack: let your agents edit and store their own Python code as a form of memory. It tests the limits of self-modification and persistent state without relying on massive context windows.

Tools

Firecrawl Gets Better at Scraping for Agents

Firecrawl’s latest update is a practical step forward for data-hungry agents. It now handles JavaScript-heavy sites and executes custom scraping functions, which means cleaner, structured inputs for your RAG or agentic loops.

Spec

MCP Spec Adds Agent-to-Agent Coordination

You need to read this if you're building a multi-agent system or orchestrator. The spec now formally defines how agents should hand off tasks and share context, giving a standard way to avoid brittle, custom glue code.

Infrastructure

AWS Bedrock Agents SDK Updates for Tool Use

Bedrock's Agents SDK just got a major update for tool use and orchestration. If you're in the AWS ecosystem, this is now the most integrated way to build and deploy production-grade agents without managing a bunch of infra yourself.

News

Groq Unveils New Inference Pricing Tiers

Groq just dropped its new pricing tiers for ultra-fast inference. The LPU machine is becoming a real option for latency-sensitive agent tasks, but you need to model your costs carefully against the speed gains.