Arga raised $10M to reset Salesforce so agents can be tested 10,000 times

Six months ago the demo was the story. Someone typed a sentence, a working app streamed across the screen, and the room lit up. Now the same teams are learning what happens when that demo meets real traffic. The gap between a good prompt and a shippable product turned out to be full of work nobody wanted to do. Let me catch you up.

The deep cut

• The invisible 80 percent is the product. The vibe-coded demo ships the interface; the parsing, access control, and evals are what survive real users.
• A model checks its work, not its premises. Across nine builds, 35 of 63 defects surfaced only when a domain expert asked the right question.
• Reset and rerun beats one lucky pass. Arga clones Salesforce so agents can be tested thousands of times; five runs catches what one hides.

The demo that lies to you

The trap is a polished screen. Modern coding tools produce something that looks like finished enterprise software: dark mode, smooth loading spinners, clean typography. Leadership sees that and assumes the engineering is 95 percent done. One breakdown from Towards AI calls this the "High-Fidelity Trap," and lists what the demo skips: schema parsing, access control, token economics, guardrails, evals, circuit breakers.

The failures are concrete. A user pastes a 40-page scanned contract and crashes the context window. A naive search pulls up executive comp memos for an intern. The invoice spikes by $14,000 because the app resends the full conversation history every turn. None of that shows up in the Friday demo. All of it shows up Monday.

So treat the vibe-coded prototype as a living spec, not a release. Use it to validate what the interaction should feel like, then hand engineering the real requirements: the boundaries that protect data, margins, and uptime.

The stuff under the interface

If the demo is 20 percent, the other 80 is architecture. Retrieval is the clearest example. A prototype ingests documents, makes embeddings, retrieves a few chunks, done. Dave R argues that flow proves the idea and nothing more. Production retrieval needs hybrid keyword-plus-vector search, graph retrieval for questions that span sources, re-ranking, and metadata access controls so cosine similarity does not leak documents across roles.

The same lesson runs through agent design. A rundown of seven agent concepts makes the point plainly: what makes something an agent is the harness around the model, not the model. The system prompt, tools, sandbox, memory, the reason-act-observe loop, guardrails, and observability. As raw model ability levels off, your effort moves to controlling what the model can see, do, and remember.

The part the model can't do for you

Here is the failure that destroys the most value, and it is not hallucination. It is output that looks pristine, passes every internal check, and is wrong about the real world. A case study on enterprise builds walks through a battery-scheduling model that assumed it won its full reserve capacity every hour and sold the same block into three products at once. The code matched a validated run, so the team signed off. A question about market structure caught it, not a bug check.

The numbers make the case. Across sixty-three defects logged over nine builds, thirty-five surfaced from domain challenge and twenty-three from adversarial review by someone who knew where to look. Two came from automated checking. A model can check its work. It cannot check its premises. This is also why that MIT NANDA figure, 95 percent of enterprise GenAI pilots delivering zero measurable return, matters: too much effort sits where the ceiling was always low.

Cheap generation does not reduce the need for experts. It raises it. When your build cycle drops from three weeks to an afternoon, you run the loop forty times instead of four, and each loop burns judgment at both ends.

Building the room where agents get tested

The reason coding agents got good fast is that code has tools to deploy, reverse, and test changes cheaply. Business software does not. You cannot reset Salesforce and rerun the same scenario ten thousand times. Arga Labs raised $10 million to fix exactly this, building full digital twins of programs like Salesforce and Workday, permission systems and webhooks intact, so agents can be trained and reset at scale. General Catalyst's Yuri Sagalov puts it flatly: a repeatable sandbox is more important with agents than it ever was with humans.

Testing discipline scales down to your own repo too. Towards AI's coding playbook offers a simple rule for stochastic systems: run an important test five times, not once. A test that passes 90 percent of the time has a 41 percent chance of failing at least once across five runs. One green check hides that.

The prompt gets shorter as the system gets smarter

The other half of this shift is context. Good agent work is less about clever prompts and more about deciding what the agent knows and when. Wyndo describes moving from a giant master prompt to a short "what do you think about this idea?" because the system already holds his audience, voice, and standards in project files. The prompt only carries what changed today.

That maps onto how disciplined teams already run coding agents. KDnuggets lists ten rules, and the through-line is engineering discipline: start with a spec, not "build the dashboard"; put persistent rules in an AGENTS.md file; make tests the contract. One caution worth flagging: a study of 100 popular repos found lint leakage in 62 percent of instruction files and context bloat in 42 percent. More context is not better context.

Three questions for your team

• When we show a demo to leadership, are we clear about which of the six or seven production boundaries we have actually built, or are we selling the 20 percent as done?
• Where in our current builds is a model checking its own work, and who is assigned to challenge the premises instead of the code?
• For the agent workflows we plan to scale, do we have a repeatable sandbox to reset and rerun, and are we running important tests five times before we trust the result?