Six engineers finished an 18-month re-architecture project in 76 days. That single data point from the AWS Kiro team, reported by VentureBeat, is why enterprises stopped asking “which AI coding assistant should we buy” and started asking “how do we run dozens of agents at once without them stepping on each other.” The answer taking shape in 2026 is the agent factory: a coordination layer that assigns work to specialized agents, checks their output, and keeps the whole pipeline auditable. This piece looks at who is building that layer, what actually moves the needle on delivery speed, and where the model still breaks down.
The Orchestration Layer, Not the Agent, Is the Product
For most of 2024 and 2025, the pitch was a better individual agent. That framing changed at GTC 2026, when Nvidia brought 17 enterprise software vendors onto a shared agent platform it builds and maintains. Adobe, Salesforce, SAP, ServiceNow, Siemens, CrowdStrike, Atlassian, Palantir, Red Hat and Cisco are among the companies that signed on, according to VentureBeat’s coverage. GitHub took a narrower but similarly telling step with the /fleet command in Copilot CLI: an orchestrator that breaks a task into independent work items, dispatches sub-agents to handle them in parallel, and reconciles the results before handing them back to a developer, as GitHub’s engineering blog describes.
Anthropic went further and folded orchestration directly into its Claude Managed Agents platform, alongside two other pieces enterprises had been building themselves: persistent memory and built-in evaluation. Netflix now runs that orchestration layer to process logs from hundreds of builds at once, and legal AI company Harvey reported roughly a 6x jump in task completion after adopting the memory feature, per VentureBeat. The common thread across Nvidia, GitHub and Anthropic: none of them are selling a smarter model anymore. They are selling the scaffolding that keeps many models productive together.
What Actually Compresses Delivery Timelines
Running more agents in parallel does not, by itself, ship software faster. It just produces more code to review. The teams reporting real timeline compression paired orchestration with spec-driven development, where agents work from a structured specification of what “correct” means before they write a line of code. That is how the Kiro team cut its own feature-build cycle from two weeks to two days, and how the AWS rearchitecture project originally scoped for 30 developers shipped with six, per the same VentureBeat report. The specification becomes the contract that lets a dozen agents work on different parts of a system without drifting from the intended architecture.
GitHub’s own documentation of /fleet is candid about the limits of parallelism without that contract. Agents share a filesystem but cannot talk to each other directly, there is no file locking, and overlapping writes can silently overwrite one another’s changes. That is precisely the failure mode spec-driven development is meant to prevent: dependencies and boundaries get declared up front, before any agent starts writing.
Enterprises Are Now Backing This With Real Capital
Factory, a startup that builds AI “Droids” for enterprise engineering teams, raised $150 million in a Series B led by Khosla Ventures, with Sequoia, Insight Partners and Blackstone also participating, reaching a $1.5 billion valuation, TechCrunch reported in April. Its clients include Morgan Stanley, Ernst & Young and Palo Alto Networks, and its differentiator is model flexibility: switching between foundation models such as Claude or DeepSeek depending on the task rather than locking into one vendor. That funding round is a useful signal on its own. Investors are no longer betting on a single breakthrough model; they are betting on the company that can route work across many of them reliably.
The Governance Gap Nobody Has Closed Yet
None of this is as settled as the funding numbers suggest. Early orchestration tools, including Salesforce MuleSoft, UiPath Maestro and IBM watsonx Orchestrate, are described by VentureBeat as “phase one”: dashboards that let IT leaders see what agents did after the fact, not systems that enforce what they are allowed to do. The question enterprises are now asking internally has shifted from “can our agents do the work” to “are our agents playing well together,” and most available tooling still cannot answer that with confidence.
There is also a lock-in question building underneath the convenience. Anthropic’s Claude Managed Agents stores session state, memory and execution graphs in a database Anthropic controls. That is useful for teams that want orchestration handled for them, and risky for teams that later want to switch providers without rebuilding their agent workflows from scratch. Before committing an engineering org to one factory model, a few questions are worth asking directly:
- Does the orchestration layer enforce a specification before agents write code, or only review output after the fact?
- Who owns the session state, memory and execution history: your infrastructure or the vendor’s?
- Can the platform route work across multiple foundation models, or does it lock you into one?
- What happens when two agents modify the same file? Is there file locking, or just a shared filesystem and good intentions?
Where This Leaves Engineering Leaders
The agent factory model works when three pieces are in place together: a specification agents can be held to, an orchestrator that can actually enforce boundaries between parallel work, and a clear answer on who owns the resulting state. Nvidia, GitHub and Anthropic are each building pieces of that stack. Factory’s funding shows investors expect a standalone company to win the coordination layer outright. None of it replaces engineering judgment; it changes where that judgment gets applied, from writing individual functions to writing the specifications and boundaries multiple agents will work inside. Teams that skip that step are the ones still finding overwritten files in their pull requests.