The Short Answer
Goldman Sachs runs hundreds of AI coding agents alongside roughly 12,000 engineers in production, and the durable advantage is the harness — evaluation, routing, governance, audit — not the model. ibl.ai is the agentic AI platform where you own all the code and the data, self-hosted inside your own perimeter, model-agnostic across any LLM, and usage-based with no per-seat pricing, so you can deploy anywhere.
Every bank can license the same frontier models. None of them can license your evaluation suite, your routing rules, or your audit trail.
That asymmetry is why the harness is the asset — and why renting it from a vendor undoes the point.
What exactly did Goldman Sachs deploy, and at what scale?
Goldman Sachs has deployed hundreds of AI coding agents working alongside its 12,000 human engineers — writing production code, not running in sandboxed demos.
This is one of the clearest signals yet that agentic AI has crossed from proof-of-concept into genuine operational deployment at enterprise scale.
The agent is Devin, built by Cognition, rolled out across the bank's technology division. CIO Marco Argenti has described 3–4x the productivity of the bank's previous generation of AI tooling, with sharply reduced development and vulnerability-fix time.
The stated trajectory is hundreds of agents scaling toward thousands. In 2026 the bank extended the program, adopting Anthropic's Claude for work spanning trades and transactions as well as client vetting and onboarding.
The second-order effects are already being priced. Analysts estimate up to 200,000 U.S. banking roles could be affected, many of them junior engineering positions.
Goldman is notable less for being early than for being unhedged: real production tasks, on the critical path, rather than an isolated innovation-lab pilot.
Why is the harness, not the model, the actual moat?
The moat isn't the model. Every engineering team at this scale has access to the same foundation models.
The moat is the harness: the infrastructure layer that wraps the model with evals, routing, access controls, and auditability.
Organizations that build that layer well don't just deploy AI faster — they deploy it safely enough that deployment becomes a repeatable process rather than a one-time gamble.
Andrej Karpathy framed the same transition at Sequoia's AI Ascent 2026, describing a shift from "vibe coding" to agentic engineering, and naming December 2025 as the inflection point when agentic coding became genuinely reliable.
His distinction is the useful one: vibe coding raises the floor on what anyone can build, while agentic engineering preserves the quality bar professional software requires. Production work lives entirely in the second category.
The demos are easy. Running AI agents where mistakes have financial consequences requires everything the demos never show.
What does a production agent harness actually contain?
A harness is not one product. It is five capabilities that have to exist before an agent is allowed near a repository that ships to customers.
| Layer | What it answers | Failure if missing |
|---|---|---|
| Evaluation | Is this agent better than last week's? | Regressions ship undetected |
| Routing | Which model handles which task? | Frontier prices on trivial work |
| Guardrails | What may the agent never do? | Unbounded blast radius |
| Audit | What happened, and on whose authority? | No answer for the regulator |
| Memory | What does the agent know about us? | Every task starts from zero |
Each row encodes something specific to your organization — your definition of done, your risk tolerance, your escalation path. That is precisely the part that cannot be bought off a price list.
What changes when AI agents write production code at a bank?
The governance question moves from theoretical to operational. A code review by a human is an accountability record; a merge by an agent needs an equivalent one, or the audit trail has a hole in it.
Three things change immediately. Attribution has to be explicit, so every change identifies the agent, the model version, and the human who authorized the run.
Rollback has to be cheap, because an agent that opens fifty pull requests can introduce a subtle defect fifty times before anyone notices.
And model versions become a compliance artifact. When a provider silently updates a model behind an API, a bank has to know that the system which passed review in March is not the system running in August.
Owning the harness is what makes all three tractable. If the routing, logging, and evaluation live in a vendor's control plane, your answer to the regulator is a support ticket.
Who should own the harness when agents run in production?
You should — because the harness encodes your institution's judgment, and renting it means renting the one thing you actually built. With ibl.ai you own all the code and the data.
The full source runs under a perpetual license inside your own perimeter. The platform is model-agnostic, so Claude, GPT, Gemini, Llama, Command, or your own fine-tune are routing choices rather than architectural commitments.
Billing is usage-based with no per-seat pricing, which matters when the workload is agents rather than people — headcount stops being a proxy for anything. Deploy anywhere: your cloud, your VPC, on-premise, GovCloud, or fully air-gapped.
More than 1.6M users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
Goldman's deployment is the proof that this era has started. The question it puts to every other institution is not which model to license, but who will own the layer that makes the model safe to use.