Five agent frameworks ship —
and WorkOS hands agents an /auth.md.

Pydantic AI cut v2.0.0b1 on Thursday and pushed b2/b3 by Saturday, alongside the steady v1.100 → v1.102 cadence on the stable line. The v2 pitch is structural, not cosmetic: "harness-first design, with capabilities as a core primitive" — the agent loop, tool surface and provider adapters get reshaped around a single composable harness object, with the assumption that you bring (and version) your own. Pending-message queues (ctx.enqueue / agent_run.enqueue) and MCP background tasks landed in the same week on both branches, so the v2 betas inherit them without migration churn. The framework that built its reputation on "just the agent loop, properly typed" is finally publishing what it learned from a year of users bolting full agents on top — v2 is the version where the harness is the product.

Two days after Python 1.5.0, Microsoft Agent Framework cut 1.6.0 with two things that matter: a built-in Shell tool that can execute locally or against a Docker container, and an experimental hosted-tool factory for Foundry. The breaking change worth flagging is operational — instrumentation is now on by default across core and Foundry packages, so every span, retry and tool call lands in OpenTelemetry without an opt-in. The right default, but it will surface in everyone's existing dashboards on upgrade. .NET 1.6.1 from May 14 added Hyperlight execution and OpenTelemetryAgent auto-wiring; the two languages are now within one minor version of each other. Microsoft is signalling that the framework is the deployment story, not just the building block.

Codex shipped its stable build of the week on Thursday as v0.133, then pushed three v0.134 alphas (.1, .2, .3) over the next 48 hours — the cadence Codex maintainers settled into the moment Workspace Agents landed in ChatGPT. v0.133's headline change: Goal mode is now on by default and backed by dedicated storage, so a Codex run owns the lifecycle of its own goal across resumes rather than re-deriving it from terminal scrollback. Plugin discovery now ships marketplace-aware list output, the Python SDK got first-class auth in v0.132, and exec resume started accepting an output-schema for structured JSON. The Workspace-Agents → CLI loop tightens every week — the CLI is where the production engineering still happens.

Four days after ADK 2.0's GA flipped the framework from a hierarchical executor to a graph engine, v2.1.0 added the most visible application of that engine: a data agent whose chart-generation and artifact-loading capabilities are first-class graph nodes, plus the ability to spin sandboxes from templates. The same release window saw v1.34.1 patch the legacy 1.x line for Gemini 3.1 grounding-metadata bugs — Google is running 1.x and 2.x in parallel during the GA cooling-off period, which is the polite version of "don't migrate yet if you can wait." The signal worth tracking: ADK's reference applications are now full vertical agents, not Hello-World loops. The framework is shipping its own opinionated agents on top of itself.

AWS's Strands SDK pushed v1.41.0 with a MultiAgentPlugin for orchestrators — filling exactly the same gap Anthropic's Managed Agents, Microsoft Agent Framework and Google ADK addressed in their May releases — plus TTL support for Bedrock prompt caching and a Starlette 1.x bump. Weekly cadence has held since v1.37 in April. The compounding effect across this section is striking: five active agent frameworks landed essentially the same primitive set (multi-agent orchestration, durable state, observability hooks, hosted-tool factories) in seven days. There is now a converged API surface for what "an agent framework" exposes; competition shifts to runtime, integrations, and price.

WorkOS founder Michael Grinich announced auth.md on Thursday: a discoverable Markdown file (typically https://yourapp.com/auth.md) plus three HTTP endpoints (/agent-auth, /agent-auth/claim, /agent-auth/claim/complete) that let an agent register itself for a service on behalf of a user. It composes existing OAuth pieces — RFC 9728 Protected Resource Metadata and ID-JAG identity assertions — rather than inventing new crypto, and ships three flows: trusted-provider (OpenAI, Anthropic, Cursor mint an assertion), user-claimed OTP, and anonymous-then-claim. Cloudflare and Firecrawl are the first publishing partners. The unsubtle point: every app on the web already has a /robots.txt for crawlers and a /.well-known/openid-configuration for humans-via-OIDC; auth.md is the equivalent for agents — and it's deliberately not tied to WorkOS infrastructure. Whether or not this exact spec wins, the category is now contested.

GitHub added Claude Haiku 4.5 and GPT-5.4-mini to the Copilot cloud agent's model picker at a 0.33× price multiplier — one-third the cost of the standard tier, for the simple-task end of the queue. Anthropic's own benchmarks put Haiku 4.5 at Sonnet-4-class coding quality at roughly twice the speed; the practical effect is that the cost of having an agent triage a backlog, run a routine refactor or generate boilerplate just collapsed. The model-tier story has finally arrived for delegated work, not just chat: pick the cheap fast model for the issue-comment queue, save the big model for the architecture call. Expect every coding-agent product to ship the same toggle inside the month.

Thoma-Bravo-owned Coupa (taken private at $8B in 2023) bought Tonkean — the no-code workflow-orchestration startup last priced at $300M in 2021 — to wire agentic intake into a spend network spanning 3,500+ buyers and roughly 10M suppliers. Tonkean brings 250+ native connectors, a natural-language request surface, and a multi-agent execution framework; Coupa is using it to power "Agentic-as-a-Service" across procurement, sourcing and trade workflows. It's Coupa's fourth agentic acquisition in eight months (after Cirtuo, Scoutbee and Rossum); terms weren't disclosed. The PE- owned-incumbent-buys-agentic-orchestration playbook is real now, and procurement was always going to be one of the first vertical SaaS categories to get rolled up under it.

A JavaScript / MIT proxy from jianshuo (~235 stars) that sits between your coding agent and the model API and renders the full request graph in a local web dashboard: system prompts, message history, token counts, cache statistics, cost estimates and a step-by-step agent-loop visualisation. Works with Claude Code, Codex, DeepSeek, Kimi, Ollama, OpenRouter and the rest — the same lane previously-covered tools cover for memory and billing, but pointed squarely at the "what did the agent see, and what did it decide" question. The agent-observability category is fragmenting into proxy / instrumentation / billing slices; ccglass is the proxy one, and the cleanest open implementation we've seen.

Fabio Akita shipped a Rust / MIT project (~153 stars) that captures every prompt, tool call and decision an agent makes into a shared markdown wiki backed by a git repo — no vector DB, just full-text search and per-project isolation. When a session ends the wiki is rewritten into a coherent narrative and prepended to the next agent's context, so Claude Code, Codex and OpenCode can pick up where each other left off. It's a direct rebuttal of the "every agent gets its own memory store" assumption — and a quiet bet that the right primitive for cross-vendor agent continuity is shared markdown, not shared embeddings. Pair with juice below for the inverse-constraint half of the same thesis.

A small TypeScript MCP server from Alvin Unreal that stores the inverse of normal agent memory: not what worked, but what to never do again. When a user corrects the agent ("don't use var, use let"), that constraint lands in juice; on the next task only the relevant constraints are surfaced to the model, so the negative store doesn't bloat context the way positive-only memory layers do. Web UI, HTTP API, plugs into OpenCode, Claude Code and Cursor. Tiny repo, unusually sharp concept — the agent-memory category has been pursuing "remember everything useful" and missing that "remember only what to avoid" might be the more durable primitive.

An Apache-2.0 / MIT dual-licensed skill collection from Klotzkette (~274 stars) covering the full sweep of German law — Arbeitsrecht (labour), Gesellschaftsrecht (corporate), Insolvenzrecht, Datenschutz, Verfahrensrecht — across 99 plugins and over 1,500 individual skills, with profiles for all 24 Fachanwaltschaften and 44 realistic test-case fixtures. The README's compliance section is the interesting bit: explicit handling of §§203/204 StGB confidentiality, DSGVO and the EU AI Act, plus pointed caveats about professional ethics. It's the most ambitious vertical Claude-skill project we've seen and a useful counterpoint to Anthropic's English-first Claude-for-Legal launch — non-common-law jurisdictions are going to build their own.

From the VILA Lab at MBZUAI: a Python agent (~237 stars) that takes a reference figure from a paper and your dataset, runs an iterative Drawer–Reviewer loop, and returns an editable matplotlib script plus a camera-ready PDF that matches the original's visual style. Ships with a 139-figure reference gallery across 25 chart types, supports 3D plots, and packages as a skill for Claude Code and Codex. Pairs with new chart-style and trading-knowledge skills landing the same week to show the same pattern: the long tail of vertical agent work is shipping as Claude/Codex skills rather than standalone products. That's the right surface — but only because skills are finally a real distribution channel.