Anthropic engineers embedded at Goldman for six months, working alongside the bank's tech team to co-develop systems. The collaboration went beyond implementation or configuration—it required building production infrastructure together.
This deployment pattern exists because the gap between model capability and production reliability is real and wide—production deployment in high-stakes environments requires bridging the distance between what models can do and what production systems demand.
"Digital co-worker" means something specific operationally. Agents validate client data, reconcile trade discrepancies, generate regulatory filings. These are multi-step workflows that process enormous data volumes, cross-reference regulatory requirements, flag exceptions, maintain audit trails.
The complexity lives in orchestrating reliable execution across steps where errors compound, where regulatory scrutiny requires explainability, where business continuity depends on predictable behavior.
Building that reliability requires engineers who understand both the model's behavior and the domain's requirements, working together to build infrastructure that handles the gap between capability and reliability.
The agents use several techniques to bridge this gap:
- Grounding techniques ensure agents reason using Goldman's actual data rather than general training knowledge—the difference between knowing regulatory frameworks in theory and applying Goldman's specific policies correctly
- Chain-of-verification prompting adds layers where models cross-check their outputs before presenting them—catching errors before they propagate through multi-step workflows
- Risk management approaches design systems around the model so errors get caught before they matter—building containment so failures don't cascade
Architecture work. Understanding AI behavior deeply enough to know where models fail, and understanding production requirements deeply enough to build systems that handle those failures gracefully. Embedded engineers bridge the gap between what's technically possible and what's operationally reliable.
The timeline reveals something about what production readiness actually requires. Six months of embedded engineering to reach a point where Argenti says they'll launch "soon" without committing to a date.
"We'll launch soon."
— Marco Argenti (no specific date provided)
That vagueness reflects recognition that production readiness in high-stakes domains requires proving reliability. You're stress-testing against edge cases. Validating in scenarios where mistakes trigger regulatory consequences. Building confidence that systems behave predictably when stakes are high. That takes however long it takes—you can't compress it by declaring victory early.
This deployment model has a shelf life. Co-development makes sense when you're figuring out how to deploy agents in environments where the stakes justify the investment and the patterns aren't yet established. But it doesn't scale. You can't embed engineers at every enterprise for six months.
The path forward requires either productizing what co-development teaches—turning custom architecture into reusable patterns—or developing infrastructure that makes deployment reliable enough that embedded engineering becomes unnecessary.
The next deployment timeline tells you where the technology is heading. If it takes three months instead of six, that's progress toward productization. If Goldman can deploy agents in new domains without Anthropic engineers on-site, that signals infrastructure maturity. If other enterprises can adopt similar patterns without custom co-development, the lessons are becoming reusable enough to work at scale.
When deployment timelines compress from six months to three to six weeks, you're seeing evidence that the infrastructure patterns needed for reliable deployment are becoming established. When vendors can deliver production-ready systems without embedding engineers, that's a maturity signal that matters.
The co-development model represents a phase that signals where we are on the maturity curve. Its existence tells you that getting agents to work reliably in high-stakes environments requires building infrastructure case by case, with deep expertise on both sides working together to bridge the gap.
When the patterns and infrastructure needed for reliable deployment become established enough to work without custom engineering, enterprise AI will look different. Deployment will happen through infrastructure you can configure rather than co-development projects you must build. We're watching that transition happen, one embedded engineering project at a time.
Right now, the deployment model itself—how Goldman had to build this—tells you where enterprise AI actually is.