Honest assessment

The advantage is structural

A comparison run against the two closest competing agent gateways found neutron’s advantage is real, and that both of those projects reject a governance plane by written charter rather than because they have not got round to it.

One states that all callers are equally trusted and that its approval prompt is explicitly not a security boundary. The other’s roadmap lists agent hierarchies among the things it will not merge.

Nobody else in the comparison set has all of:

multi-user human roles
separation of duties, enforced by the policy engine
agents managing agents under a non-escalation rule
many tenants provisioned from one image

The counterweight is equally real

Those same competitors connect to far more services and channels than neutron does — one supports roughly thirty providers and thirty channels, the other closer to fifty-seven providers and twenty-five channels. One has a self-improving skill and memory loop that neutron has no answer to.

Neutron is narrow and deep where they are broad and shallow. That is a positioning, not a victory.

Neutron’s conversation loop is a third-party agent SDK. “More capable than the underlying coding agent” is a category error — the claim is about the platform layer around it, not the model loop inside it.

The five gaps

Gap	Why it matters now
No tenant column	Separation is one pod and one database per customer. Safe, but adding a customer costs infrastructure rather than a database row.
Single replica	Some state lives in the running pod, so you cannot simply run two copies for redundancy yet.
No control plane	Signup, billing, provisioning and fleet-wide upgrades are not built. Someone does them by hand today.
No metrics export	Costs and audit are recorded inside the product but not shipped to standard monitoring.
Shared credentials	On a developer-facing instance, one person’s third-party account can be spent by everyone using it. A known open issue.

These matter more than they used to, because the go-to-market decision moved to hosted-service-first. Self-hosting is still supported; it is simply no longer the assumption.

The upgrade blocker that used to sit underneath all of this — needing to recreate the database to move versions — was cleared when incremental migrations shipped.

What is genuinely settled

The permission engine is extracted, published, and consumed back by the platform. That round trip is the proof it is a real boundary rather than an aspiration.
The coding controller produces reviewable merge requests on both GitLab and GitHub, and refuses to merge them.
Two instances run the same image with different configuration, which is the multi-instance claim demonstrated rather than asserted.