Methodology
Marks you can check.
Capability claims need supporting documentation or evidence. When we cannot find support for a criterion, we say “not claimed” rather than guess.
The six criteria
Approval gates
Can a person or policy hold an action until it is approved? Sensitive or irreversible actions can wait for an explicit human or policy decision before they run, and the wait survives longer than a single request.
Restart recovery
Does work resume after a crash or redeploy without repeating finished steps? Progress is persisted outside the process. After a crash, restart or redeploy, work continues from the last completed step instead of starting over or silently dropping.
Evidence trail
Can you inspect what ran, with what inputs, and who approved it? Actions, inputs, outputs and decisions are recorded in a form an operator or auditor can review after the fact.
Kill switch
Can an operator withdraw permission or stop further actions? An operator can revoke, cancel or terminate an agent run so that further actions do not start. Note whether an action already running is interrupted.
Policy controls
Do declared rules decide which tools and actions are allowed? Tool access is scoped by declared rules (allowlists, RBAC, CEL, Cedar or similar) that are enforced outside the model, not by prompt instructions.
Sandbox isolation
Does code or tool execution run inside an isolated boundary? Execution happens inside a container, microVM or similar boundary that limits what a compromised or mistaken action can reach.
How marks are assigned
- Built in — the vendor documents the capability as a feature of the product.
- Documented pattern — the vendor documents primitives or a recipe that achieves it, but you assemble it. Example: a workflow that pauses until a signal, used as an approval gate.
- Not claimed — we could not find documentation for it, or have not reviewed it yet. This is not a claim that the capability is impossible.
Each listing shows the date it was last reviewed and the sources behind its marks.
Verification levels
- From public docs — marks come only from the linked documentation.
- Evidence linked — the vendor publishes test evidence (CI runs, evaluation reports, release checksums) that we link. We have not re-run it.
- Not yet verified — supporting checks have not been completed for this directory.
- Verified — we installed the product and ran a documented check for each criterion marked “Built in”, and we publish the evidence with the date and version tested. A product changes after it is verified; the badge applies to the version named.
Ownership and money
Durable Agent Directory is run by NextEleven. Products built by NextEleven — Mothership, DeadBolt, Harness, Witness and others — carry a “By NextEleven” label and are placed first. We hold them to the same marking rules, and their Verified badges require the same published evidence.
Vendors can ask about future placement. No paid placement or verification offer is open here. Any future placement would be disclosed and would never change criteria marks. Outbound links carry directory attribution; basic third-party links are nofollow and paid placements, if introduced, would be marked sponsored.
Corrections
If a mark is wrong or out of date, send a correction with a documentation link. We update listings when the documentation supports the change.