Agents: build fast, add memory

An AI agent is a language model wrapped in tools and rules so it can plan, call software, and finish multi-step jobs instead of answering in one shot. OpenAI’s current Agents SDK says agents “plan, call tools, collaborate across specialists, and keep enough state to complete multi-step work.” (developers.openai.com) That tooling has gotten easier to reach. OpenAI’s Python Agents SDK repository shows 24,300 stars on GitHub as of April 22, 2026, and the docs pitch a “quickstart” as the shortest path to a working integration. (github.com, developers.openai.com) The newer pitch is not one agent doing everything, but a small team with separate jobs. Microsoft’s Azure Essentials Show on Oct. 23, 2025 laid out sequential, concurrent, group chat, and hand-off patterns for specialized agents working together. (learn.microsoft.com) OpenAI’s own cookbook uses the same structure in a finance example published May 28, 2025: a Portfolio Manager agent delegates work to Macro, Fundamental, and Quantitative specialists. The guide contrasts that “agent as tool” setup with handoffs, where one agent passes control to another mid-task. (developers.openai.com) Anthropic has made a similar case from production use. In a June 13, 2025 engineering post, the company said its Research feature uses a planning agent that spawns parallel subagents to search different directions at once, then compresses their findings back into one answer. (anthropic.com) The reason builders keep adding more structure is that long jobs break simple prompts. Anthropic said multi-agent systems introduce coordination, evaluation, and reliability problems, while OpenAI’s docs point developers toward guardrails, human review, tracing, and explicit orchestration when one specialist is not enough. (anthropic.com, developers.openai.com) Memory is the other piece moving from optional to standard. OpenAI’s sessions docs say the SDK can automatically maintain conversation history across runs, and Anthropic’s Managed Agents docs describe stateful sessions with persistent file systems and conversation history across interactions. (openai.github.io, platform.claude.com) That is still short-term memory unless developers add a workspace the agent can revisit. Obsidian’s own materials describe the app as a knowledge base built on local Markdown files, which is why creators and developers keep pairing agents with note vaults, databases, and shared folders instead of relying on raw chat history alone. (github.com, obsidian.md) The practical model in 2026 is simple: start with one agent, split repeatable subtasks into specialists, and give the system a place to store what it learns between runs. The software companies now ship the scaffolding; the remaining work is deciding what your agent should remember, and what it should forget. (developers.openai.com, platform.claude.com)