Independent AI intelligence Two editions daily · ET
FervorAI
Menu

AI Trending Briefing · July 23, 2026 · morning edition

The same week vendors raced to ship enterprise agent-deployment platforms, the plumbing beneath them (MCP) was rebuilt for stateless scale and hardened auth, and a wave of runtime governance tools arrived to watch what those agents actually do.

OpenAI PresenceMCP 2026-07-28 SpecAlterion DracoClaude CoworkGemini 3.6 FlashAxonius MCP Servermcpagent-infrastructureagent-securityclaude-codecodexfrontier-models

Trending AI Briefing: Thursday, July 23, 2026 (morning ET)

Agents stopped being a demo this week and started being an operations problem. Three things landed inside 48 hours that only make sense together: a platform to deploy agents under corporate policy, a protocol rewrite that lets those agents scale like ordinary web services, and a governance layer that sits in the request path to watch what they do. OpenAI Presence, the MCP 2026-07-28 spec, and Alterion's Draco are all answers to the same question, which is no longer "can an agent do the task" but "who is accountable when it does."

What's hottest in AI news right now

OpenAI Presence shipped July 22 as an enterprise product for deploying and managing agents across customer-facing and internal workflows. The framing is the story: each deployment starts with one narrow job, like resolving a billing issue or an IT service request, and the agent gets only the knowledge and system access that job needs. The company sets the policies for what the agent can do, when it needs a human to approve, and when a person takes over. Presence is not self-serve. It ships through a limited GA program with OpenAI's Forward Deployed Engineers and select systems integrators running the deployments, which tells you exactly who OpenAI thinks the buyer is (VentureBeat, OpenAI newsroom).

The MCP 2026-07-28 specification is the plumbing underneath all of this, and it is the biggest revision since the protocol launched. Two proposals do the heavy lifting: SEP-2575 removes the initialize handshake, and SEP-2567 removes the Mcp-Session-Id header and the protocol-level session entirely. Client info, protocol version, and capabilities now ride in a _meta field on every request, so each call is self-describing and remote servers can run behind a plain round-robin load balancer with no session affinity. Session-based interaction gets replaced by Multi Round-Trip Requests, where a server that needs user input mid-call returns an InputRequiredResult with a requestState blob. The release candidate locked May 21, with the final spec due July 28 (MCP blog, 4sysops breakdown).

Alterion Draco launched July 16 as a runtime control plane that sits between agents and enterprise infrastructure. It observes prompts, actions, and payloads from production agents and enforces programmable guardrails in real time, without touching agent code. Alterion's pitch names the gap directly: APM and security tooling capture logs and metrics but not agent intent, which leaves compliance holes against SOC 2, ISO 42001, and the EU AI Act. The founders are a former McKinsey partner and a long-time Google engineering VP, which is a fair signal of who they expect to sell to (PR Newswire, Alterion).

Claude Cowork is moving to the cloud. Anthropic said it is taking its desktop agent off the single-device model so sessions and files follow you across mobile and web, with background work, scheduled tasks, shared chats and projects, and approvals you can grant from your phone. It is the same accountability theme from the consumer side: the agent keeps working when you are not watching, so the interesting design question becomes how you approve and audit what it did (NBC News, Anthropic newsroom).

Google DeepMind released three Gemini models July 21, and skipped the one everyone wanted. Gemini 3.6 Flash is the new workhorse, with better coding and multimodal work while cutting token use by up to 17%, which makes it cheaper than 3.5 Flash. There is also 3.5 Flash-Lite for cost floors and, notably, Gemini 3.5 Flash Cyber, fine-tuned to find and fix security vulnerabilities and restricted to governments and trusted partners under a limited pilot. Gemini 3.5 Pro still is not generally available, after an earlier mid-July target slipped (TechCrunch).

Axonius shipped an AI Agent and an MCP Server on July 21, letting security and IT teams ask asset-intelligence questions in plain English and get answers in seconds. It is a smaller launch than the others, but it is the same pattern from the security-ops corner: wrap an existing data platform in an MCP endpoint so agents can query it under existing access controls (GlobeNewswire).

New tools and features worth actually trying

Claude Code v2.1.218 (July 22) turned /code-review into a background subagent, so a review no longer floods your main conversation, and /fork now copies your session into a new background session while you keep working. Both are real quality-of-life wins for anyone running long agent sessions. The tradeoff is spend: background subagents burn tokens you do not see scroll past, which is why the same recent releases added a default 200-cap on subagent spawns and WebSearch calls per session. If you are on a metered plan, watch those numbers (Claude Code changelog).

OpenAI Codex added /import to migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, commands, and project memories in one move, plus Codex Remote is now GA from the ChatGPT mobile app against a connected Mac or Windows host. The /import alone makes trying Codex nearly free if you already live in Claude Code. The catch on Remote is that it needs a machine you own left on and connected, so it is not for people who work fully in the cloud or on a locked-down corporate laptop (Codex changelog).

Gemini 3.6 Flash is worth a look if you run high-volume agent loops where token cost dominates, since a 17% reduction compounds fast across millions of calls. Be honest about what it is, though: a workhorse Flash model, not a frontier reasoner. If your bottleneck is hard multi-step reasoning rather than throughput, this is not the model that fixes it, and 3.5 Pro still is not available to fill that gap (TechCrunch).

Alterion Draco is worth a pilot if you already have agents in production and a compliance team asking what they touched. Enforcing guardrails without changing agent code is the right design. The tradeoff is that it is one more proxy in your request path, and it is priced and built for enterprises, so a solo builder or small team gets most of the same safety from scoped API keys and good logging (Alterion).

The MCP stateless migration is worth reading before July 28 if you run any remote MCP server, because the payoff is concrete: ordinary HTTP hosting behind a normal load balancer, no sticky sessions. The tradeoff is migration work. Roots, Sampling, and Logging are deprecated, and while the spec promises a twelve-month deprecation window, anything you built on session state needs a plan now, not in a year (MCP blog).

Trending AI repos on GitHub today

Numbers below come from Trendshift's daily momentum board and reflect that ranking, not a direct GitHub star pull, so treat them as directional. Verify before you cite any figure downstream.

lidge-jun/opencodex: a universal provider proxy that lets you drive the Codex CLI, App, SDK, and Claude Code with any model (Claude, Gemini, Grok, DeepSeek, Ollama). Sitting at number three the same week Codex shipped /import says builders want out of single-vendor lock-in (Trendshift).

block/buzz: a self-hosted "hive mind communication platform" tagged as an AI-agent project, holding second place. Multi-agent coordination infrastructure keeps surfacing as its own category, separate from the coding agents (Trendshift).

microsoft/SkillOpt: a text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits and validation-gated updates, emitting a deployable best_skill.md. It is a research-backed take on the skills pattern that is eating the agent-tooling world (Trendshift).

tirth8205/code-review-graph: a local-first code intelligence graph for MCP and CLI that builds a persistent map of your codebase so AI tools read only what matters. The pitch is benchmarked context reduction on reviews and large repos, which is the real cost lever in agent coding (Trendshift).

stablyai/orca: an "agent development environment" for running a fleet of parallel coding agents with your own subscription, on desktop and mobile. Parallel-agent orchestration UIs are quietly becoming a product category (Trendshift).

pireel/pireel: a backend-free, in-browser AI video editor for talking-head video, drivable by any agent over MCP. Notable less for video and more as evidence that "drivable over MCP" is now a default feature builders advertise (Trendshift).

mattpocock/skills: a straight dump of one working engineer's .claude skills directory, near the top of the board. The signal is that curated, personal skill libraries are becoming shareable artifacts in their own right (Trendshift).

koala73/worldmonitor: today's number-one repo, a real-time global intelligence dashboard doing AI news aggregation and geopolitical monitoring, tagged AI infrastructure and local LLM. Adjacent to the core beat, but a clean example of agent-style aggregation shipped as an app (Trendshift).

What actually matters from today's signal

The clearest trend to track this week is the split between shipping agents and controlling them, and both halves moved at once. If you build, the four highest-signal areas right now are the MCP stateless migration (it changes how every remote server deploys), policy-bounded deployment platforms like Presence (the enterprise buying pattern is now visible), runtime governance planes like Draco (a new layer in the stack that did not exist six months ago), and the skills-as-artifacts wave showing up all over the trending board. Those are not four trends. They are one stack forming in real time.

Here is the counter-signal. Presence is not self-serve, Draco is built for enterprises, Gemini 3.5 Flash Cyber is restricted to governments, and the most interesting frontier model, Gemini 3.5 Pro, still is not available. The governance story is real, but so is the gate. The tooling that makes agents accountable is arriving first for large organizations with integrators and compliance teams, while the solo builders who drove agents to this point get the proxy hacks and the DIY skill dumps. Watch whether the control layer commoditizes down to individuals, or whether "safe agents" quietly becomes an enterprise-only tier.

And the risk that rides along with the trend: every governance product on this list adds a component that sees every prompt and payload your agents handle. Draco, Presence, and the MCP servers wrapping your data are all now high-value targets sitting in the request path. The same week the industry got serious about watching agents, it also built a lot of new places for a breach to matter more. That is the story worth following into next week.

Source access notes: Vendor blog homepages (anthropic.com/news, openai.com, blog.cloudflare.com, blog.langchain.com) and direct GitHub repo pages were blocked by the sandbox web_fetch provenance rule this run; all vendor and repo facts were sourced through WebSearch results and the Trendshift daily board instead. GitHub star counts are Trendshift's daily figures, not direct pulls, and are flagged as directional in the repo section.