Oracle wants your terminal

Happy Wednesday, there 👋

Yesterday we said the giants were fighting over the plumbing, and that the smart move was to stay portable. We didn’t expect the counter-evidence to arrive in 24 hours. It didn’t. The confirmation did.

Here’s the part that should make you sit up. Oracle — the most closed enterprise stack in the business, the company whose entire personality is “our tools, our runtime, our license” — spent Tuesday telling customers to build Fusion agents in VS Code, ship them through Git, and point a coding agent like OpenAI Codex or Claude Code at the work.

That’s not Oracle loosening up. That’s Oracle deciding the editor was never the thing worth owning. Watch what they kept.

The Big Thing

Oracle just told you to bring your own coding agent to its ERP

On July 14, Oracle announced an AI-native builder experience for Fusion Cloud Applications. The headline feature is the AI Studio Skill — a pro-code path into Oracle AI Agent Studio. You build agents in your own editor, run standard CLIs, use your own Git workflow, and hand the grunt work to an AI coding agent. Oracle names OpenAI Codex and Claude Code by name. No-code, low-code and pro-code now sit on one platform.

Here’s the delta, since we covered Oracle two weeks ago. Two weeks ago in Issue #26 we wrote up Oracle’s June 29 announcement of four supply-chain Fusion Agentic Apps plus Agent Studio — prebuilt agents and a place to configure them. That was the what. Today is the how: the door out of Oracle’s console and into your toolchain. Same platform, new front end — and the new front end is yours, not theirs.

Now the third-thought read. Ask what Oracle gave away versus what it did not. It gave away the harness: the editor, the CLI, the model, the coding agent. Those are yours to swap on a Tuesday afternoon. It kept Fusion’s business objects, approvals, and audit log. It kept the record of who authorized the agent, what data the agent touched, and what the agent did.

That’s not generosity. That’s a company that knows exactly which half of the stack is defensible — and it is not the half you type in. Chris Leone, EVP of Applications Development, is the exec on the release. Reporting from InfoWorld and SiliconANGLE says there’s no added cost for Fusion customers and partners; treat that as reported, not as Oracle’s own words, until you see it in your contract.

Ship it? Deploy — if you run Fusion. This is shipped product with a real path in, not a roadmap slide. Take one developer, one pro-code agent, and one repeatable back-office task you already understand end to end. Build it in your editor, land it through your Git flow, and read the audit log afterward like you’re the auditor. The build gets cheap this quarter. The proof does not.

Sources: Oracle newsroom primary (Jul 14, 2026), full release body verified. Cost detail attributed to InfoWorld/SiliconANGLE reporting.

Tour de Headlines

🧾 An operating system whose whole pitch is “show your work.” Also on Jul 14, Alation unveiled AIOS — an intelligence operating system for governed agents, with its own Agent Studio on top. Skip the OS branding for a second and look at the product: the selling point is lineage and decision traces. Not smarter answers — provable ones. Where did this agent’s context come from, how fresh was it, what chain of reasoning produced the output. That’s the same asset Oracle held back yesterday, sold as a standalone layer. If you’re piloting agents on your catalog, here’s a concrete change: write acceptance tests that assert data freshness and lineage, not just answer quality. A right answer from stale, untraceable context is a defect that passes your tests.

🔐 A program about agent identity — with nothing to install. Entrust opened its Agentic AI Trust Accelerator on Jul 14, aimed at four things: verifiable identity for humans and agents, real-time authorization, cryptographic key and certificate protection, and proof-of-action records a regulator can examine. Be clear about what this is: it’s a program, not a product. There’s no download, no SKU, no pilot to spin up this week. Entrust cites an IBM study on governance lagging AI adoption — that’s their framing of someone else’s research, so weigh it accordingly. Ship it? Watch. But note the pattern: a security vendor’s opening bid in 2026 is identity and proof, not tooling. Nobody is racing to give away that half.

🧠 The same agent, free on your laptop or rented by the month. Hermes, the agent from Nous Research, is the portability argument in one product. Run it locally for nothing, or take the hosted tier at roughly $20–200/mo. Identical agent, and the deploy choice is a line item — not a migration. That’s what “above the seam” looks like when it’s working for you. The money news: TechCrunch reports Nous is in talks for about $75M at a $1.5B valuation with Robot Ventures and USV. In talks — not closed. Read it as a signal on open-source agent demand, nothing more. One governance note before you get excited: Hermes can extend itself with new skills, so review who can publish into that skill repo before you let it self-extend inside your org.

Tool of the Day

🔎 future-agi

Open-source evals, tracing and guardrails for agents — self-hosted, Apache-2.0, free.

Yesterday we argued you should pour effort into governance and verification, because no standards war can commoditize them. future-agi is that argument as a repo you can clone this morning. Apache-2.0, self-hostable, actively maintained — the last push landed today.

Honest scale check: 1,407 stars and 390 forks. That’s a small project, not a movement. Read the code before you trust it in anything that matters, and don’t expect a big-community safety net. But it’s free and it runs on your infrastructure, which means the one thing that stops most teams — a purchase decision — isn’t in the way.

The move is small on purpose. Take the single agent you already have in production. Add tracing. Add one eval. Not a governance program, not a committee — one eval, today. You cannot govern what you cannot see, and right now most teams cannot see.

Clone the repo →

Worth a Click

• The bigger-community version of today’s tool — AgentScope (Apache-2.0, ~27.9k stars) builds on one line: run agents you can see, understand and trust. July updates add Kubernetes support, an OpenSandbox workspace, and long-term memory. If future-agi is too small for your risk tolerance, this is the same thesis with a much larger crowd stress-testing it.
• What the seam looks like in a regulated room — Abrigo announced APX, an agentic lending platform running pipeline through underwriting, closing and servicing — document collection, exception handling, QC. GA expected Q3 2026. What makes it worth two minutes isn’t the agents; it’s that policy guardrails and audit history travel with the work into a room where examiners show up. Abrigo estimates over 40% less manual labor — their number, their estimate, so treat it as a claim.

Yesterday we predicted it. Today we watched three vendors do it in one week.

Look at what shipped. Oracle gave away the harness and kept the audit log. Alation built an OS around lineage. Entrust convened a program around identity and proof-of-action. Nobody fought over editors, because the editor stopped being worth fighting over.

So the standards war isn’t resolving in a protocol. It’s resolving in a seam. Above it — editor, model, framework, the agent itself — everything commoditizes toward zero and toward portability. Hermes is free on your laptop; future-agi is Apache-2.0. Below the seam sits the boring quartet: who authorized this agent, what context did it read, what did it do, and can you prove it. That’s where incumbents are digging in, because that’s what enterprises actually buy.

Your Monday move costs nothing: instrument the one agent you already have running.

Pick your tools like rentals. Pick your evidence like it’s the mortgage.
— Ace, for The Agent Stack

You’re receiving this because you subscribed to The Agent Stack. · Unsubscribe