Harness Beats Raw Intelligence: Lessons from Oracle's New GitHub Drop

An Oracle ACE's read of the fusion-ai-studio repo: a node-by-node census of all 94 sample workflows, the evaluation harness hiding in the CLI, and what to check before your team clones it.

By the time Oracle’s 14 July live session ended, the feed had already decided the story: Chris Leone’s pro-code push, live builds within days, recaps close behind. All true, and a real milestone. But the newly published repo was hiding more than the demos let on.

The morning after, I cloned the repo instead of building an agent. Every README, all 94 sample workflows, the CLI’s full command surface, both how-to guides - with Claude reading and parsing alongside me, because 94 workflow files is not a weekend well spent by hand. The repo says more about where Oracle thinks enterprise agents are going than anything said on the Live. Some of it Oracle hasn’t said out loud at all.

The most revealing thing in it isn’t a demo. It’s a ratio: across all 94 sample workflows, Oracle builds four deterministic nodes - plain code and branching logic, no model call - for every one LLM call. Here’s what that, and the rest of the drop, tells you.

Quick summary - the headlines:

• The sample library is 83% not AI. Across 94 workflows there are 1,481 pipeline nodes; only 247 (16.7%) call an LLM, and plain CODE nodes outnumber LLM calls 420 to 247. Twenty-eight workflows - almost a third contain no LLM node at al - This is intentional design choice - every LLM node is a latency risk and a potential billable event.

• Every sample is a workflow graph. Zero .agent or .topic files ship in this drop. Read it as a steer: workflow teams are the pattern Oracle trusts inside the 60-second response window.

• The CLI ships a full evaluation harness - test generation, LLM-as-judge scoring, model sweeps, token analysis - that the install guide doesn’t cover yet. It’s the most forward-looking thing in the box, and no independent run of it has been published.

• The quarterly update replaces your folders wholesale. Only src/ survives. Fork a sample in place and the 26D uptake deletes your work.

• Codex is the documented assistant, not a requirement. The skill format is the open Agent Skills convention; the same skill runs under other coding agents unmodified.

Read on for the full breakdown.

What Actually Landed

The repo (github.com/oracle/fusion-ai-studio) was created on 26 June and went public alongside the 14 July announcement. Inside the release-26C/ folder: a VS Code extension (a .vsix inside a ZIP), the aistudio agent skill with a Node CLI inside it, two domain skills (succession management and warehouse shortages), and a sample library spanning eleven HCM modules, General Ledger insights, and three SCM modules.

The housekeeping tells its own story: UPL licence, no external pull requests accepted, one branch, no tags, publish-only commits. Development happens inside Oracle, GitHub carries the releases. This is a delivery medium for a quarterly drop, not an open-source project in any community sense - there’s no pull request to file, no maintainer to watch respond to an issue.

The functional shape is genuinely new for Fusion, though. Agent Studio artefacts - workflows, business objects, apps, tools - now exist as JSON files on your laptop: created locally, edited locally, pushed to the environment through a CLI. Configuration you can diff, review, and version. Leone framed it as the foundation for modern AI software engineering, and for once the architecture backs the sentence.

What the Sample Library Really Teaches

The demos show how fast you can build. The samples show how Oracle thinks you should build - and the single most useful thing to learn here is the number: four deterministic nodes for every LLM call.

Use it three ways:

• As a design-review shortcut. When your proposed workflow is mostly LLM nodes, the question should be - which of them a CODE or SWITCH node could replace?

• As a cost-model input. The node mix, not the use-case count, is what drives AI Units consumption - the ratio gives you a realistic baseline before anyone signs a forecast.

• As a scoping anchor. Start from the nearest sample, not a blank canvas. The median shape is ten nodes; anything sprawling past xx_worker_search (roughly 80 nodes) probably needs decomposing.

Here’s the count behind it. I had Claude parse the dataPipeline of all 94 sample workflows and count every node.

Strip the START/END scaffolding and deterministic nodes outnumber LLM calls roughly four to one. The median sample has ten nodes, two of them LLM. The largest, xx_worker_search, runs roughly 80 nodes with 13 LLM calls. And 28 of the 94 - almost a third of the reference library - contain no LLM node whatsoever.

The samples are also exclusively workflow graphs: no .agent or .topic artefacts in this release, and the README marks Workflow Teams as the recommended architecture. Read that as a steer. In the agentic-app context, where every response has to land inside the 60-second window, Oracle is showing you the pattern it trusts.

And the restraint is the craft. Every LLM node is a latency risk and a billable event. I covered the AI Units arithmetic in my pricing piece, these samples read like they were built by people who did that arithmetic too. In a field where most vendor demos maximise the AI on show, shipping a reference library this disciplined is a quietly confident move.

Hold onto this: the sample library is a style guide. The fastest way to raise your team’s first build is to make reading it part of the design phase.

The Evaluation Harness Is the Quiet Headline

The install guide walks a first-time user through setup with genuine care, and it stops there - fair enough for a first release aimed at getting people building. The CLI’s --help output is where the release shows its longer game.

Roughly 25 commands sit beyond the guide: do-generate-workflow-test, run-workflow-tests, compare-workflow-test-runs, run-optimization-sweep, analyze-token-usage, and more. Together they add up to LLM-as-judge evaluation, regression comparison across runs, per-node model overrides, and automated cost-optimisation sweeps - sitting next to a full policy lifecycle and the non-interactive login a CI pipeline needs.

Oracle showed this working at the 14 July session. The live model-downgrade demo switched two of six capabilities to a lower-tier model for the same quality at lower cost and latency. That demo maps directly onto run-optimization-sweep and get-workflow-model-override-targets. This is the part of the release that turns agent building from a craft into a discipline, and it deserves more attention than the fast-build demos you have seen so far. The honest caveat: the harness is demo-proven, not field-proven. The bundled guide doesn’t cover it, the samples ship without example tests, and I can’t find a published account of anyone running it independently - mine included, which is why it’s next on my lab list. Documentation usually catches up a release or two after the machinery lands. What matters is what you do in the meantime.

The real so-what, in three parts:

• Your definition of done changes. Once test generation and judge scoring are one CLI command away, a workflow without a test and a judge result attached is a draft, not a deliverable. Write that into your delivery standard now, ahead of the tooling becoming table stakes.

• Your evidence trail changes. When an AI assistant authored the workflow, the judge results are the review artefact your auditors - and your own QA - will actually accept. Keep them with the release record.

• Someone needs to own it. Give the harness to whoever owns test scripts on your programme today, and have them run it against one sample workflow this month. The teams that learn it early will set the standard the rest inherit.

The Quarterly Update Eats Your Folder

The second how-to in the repo covers uptake, and it is blunt: when a new release drops, you replace aiapps/ wholesale, the aistudio skill folder wholesale, every domain skill folder wholesale. The only directory that survives is src/ - the things you built yourself.

Follow the thread. The natural first move for any consultant - copy the succession-management sample, rename it, adapt it - is precisely the move the update model punishes. Copy the pattern into src/, never fork the file where it stands. And every quarter, everything your src/ work sits on shifts underneath it.

Which is exactly what regression evaluation is for. The uptake model and the evaluation harness are two halves of one story: a quarterly refresh of the base is comfortable only if you can re-run your evals against it and diff the results. The survival kit ships in the same box - the guides just haven’t joined the dots yet, so join them yourself.

The sentence to lift into your delivery methodology: re-baseline your workflow evals as a standing step of every release uptake.

Codex on the Box, Any Agent in Practice

The bundled how-to sends you to developers.openai.com to install Codex, and every demo you’ve seen runs on it. Oracle’s own press release is broader - it names OpenAI Codex and Claude Code as supported assistants - but the only install guide in the repo documents Codex alone, and that guide is what your consultants will actually follow. Worth being precise about what this does and doesn’t mean.

The skill itself is packaged in Agent Skills format - the open standard Anthropic published, since adopted across the major coding agents: a SKILL.md with prompt references and scripts under .agents/skills/. Nothing in it is Codex-specific, the CLI is plain Node, invoked by whatever coding agent reads the skill. I’ve had this exact skill running under Claude Code and GitHub Co-Pilot in my own setup. Codex is the documented path, not a dependency.

Three practical consequences:

• No assistant lock-in. Whatever coding agent your firm has already put through security review can drive this.

• A data-governance question - flag it to your CTO early. The documented flow has functional consultants signing into a third-party AI assistant and handing it local files describing your org structures, GL logic and talent data context. Someone needs to answer which enterprise agreement that session runs under before the first laptop does it in anger.

• A second meter running. Every coding-agent session against the CLI burns its own tokens, on top of whatever AI Units the workflow consumes once live. Budget for the assistant plan alongside the AI Units - agentic development bills on both sides of the keyboard.

And for those keeping score in the protocol wars: nobody abandoned MCP here. MCP remains a first-class tool type inside Agent Studio, where a running agent calls a live system. The CLI and skill live on the other side of the line, where artefacts get authored. Runtime integration and build-time authoring were never the same problem, and Oracle has just drawn the boundary where the rest of the industry has been converging on it.

Before Your Team Clones It

Credit first: the platform closes more of this than a generic security take admits. The CLI encrypts the Basic Auth password it stores in env.properties. Artefacts pushed to the server run inside Fusion’s security model, and business-object access rides the requesting user’s context. The residual list is short - which is rather the point:

• The extension is a side-loaded .vsix out of a ZIP. Enterprise endpoint policies commonly block exactly that. Getting it whitelisted is your first blocker, and it isn’t a technical one - start it before the kick-off, not after.

• Oracle’s .gitignore ignores only a single build script, and the quick-start builds inside the cloned repo. The moment a workspace goes under git - the whole direction of travel - env.properties is one careless git add -A from being committed, carrying your pod URL and username even with the password encrypted. Write the ignore file first.

• One environment per workspace. env.properties holds a single host. If test is on 26C and prod is still on 26B, that’s parallel workspaces and parallel skill versions - plan the folder layout before ten consultants invent ten conventions.

• Nothing mandates a review before a push to a live environment. do-save-workflow deploys, and the toolchain puts no gate between the coding agent and the save. For HCM that’s a conversation, for the General Ledger samples it’s an audit finding waiting to be written. Put a human diff in the path, and keep the judge results as your evidence trail.

What To Do This Week

Three moves, in order:

1. Clone the repo and run the node census on the samples nearest your own use cases. The pattern you should copy is sitting in there, counted.

2. Get the .vsix in front of whoever owns your endpoint policy. It’s the long pole, and it isn’t technical.

3. Run do-generate-workflow-test against one sample workflow and see the judge machinery with your own eyes - before anyone on your programme claims an agent is production-ready without it.

The samples answer what Oracle thinks agents should look like. The harness decides whether yours actually work. No doubt, Oracle will keep refining this and adding to it but there's already enough here to be worth the time of anyone creating, deploying or managing agents today.

Sources: the oracle/fusion-ai-studio repository, its bundled guides and CLI, and my own analysis scripts against a fresh clone. Node counts are reproducible from the repo’s release-26C sample set. Oracle product names and any screenshots remain the property of Oracle Corporation. Views mine, not my employer’s.