Gartner says 70 percent of enterprises will drop vendor-built agentic AI by 2028

Gartner predicts that by 2028, 70 percent of enterprises will abandon agentic AI systems built with vendor assistance, The Register reported on 30 September. The consultancy points to a model it calls forward-deployed engineering (FDE). A vendor embeds its own engineers with a customer and builds and deploys software for that customer's specific needs.

That approach can deliver fast early progress. It can also leave the customer unable to maintain the system once the vendor leaves, Gartner argues. FDE "success starts with getting the engagement structure right, from scope and incentives to governance, ownership, and exit," said Gartner senior director analyst Mukul Saha, according to The Register. Gartner also predicts that through 2028, fewer than 20 percent of FDE engagements will turn recurring customer requirements into features in the vendor's core product. It warns of "FDE washing," where ordinary consulting services get marketed as something more specialized. The advice: use FDE only for problems that need deep product expertise, rapid adaptation, or tight integration with the customer's operating environment.

The tooling gap vendors say they are closing

The same day Gartner's warning landed, OpenClaw announced OpenClaw Enterprise, an open source control plane for managing persistent agents in sensitive environments. The project says IT departments often ban agentic platforms outright because they lack a common security, safety and governance standard. OCE is currently pre-1.0 and developed in the open. It adds multi-tenancy, hard security boundaries and auditability across the agent lifecycle, and the project says Red Hat and OpenAI are already piloting it.

A separate open protocol published on GitHub the same day, UAI, targets identity and accountability rather than deployment. It binds an agent to an accountable owner, defines what it may do, records what it did and lets a third party verify the evidence without trusting the registry that produced it. The project is explicit that it does not claim an agent is safe, only that its actions can be made attributable. Network access is another control point. Tailscale published a walkthrough on 30 September of how Meta's Muse agent joins a user's tailnet as its own node, with grants, tags and one-time confirmations governing what it can reach. Tailscale describes the setup as a layer on top of Meta's own safeguards against prompt injection and data exfiltration.

Builders push back with speed and cost claims

Meanwhile, tooling vendors are competing on throughput. Magnitude, a YC-backed open source inference engine, says it tunes kernels on the user's own hardware so open models run up to 2x faster than llama.cpp, with 92 percent faster decode on Metal and 19 percent on CUDA. git-dedup, a drop-in git replacement for agent fleets, claims large checkouts run 6x faster and cut one developer's git storage from 35.5 GB to 11.1 GB.

Vercel's State of agent skills report, dated 25 September, says skills.sh reached more than one million listings and nearly 280 million recorded installs in seven months, with cross-industry skills accounting for 87.5 percent of installs in its classified sample. Those are registry activity figures, not proof of business value. DevNavigator, which covered the report on 30 September, notes that instructions can also encode outdated assumptions or individual bias, so subject-matter review still matters.