Control & trust

Which AI agent platforms require human approval before an agent acts?

Of 955 vendors, 483 document full human oversight: approval steps, consent checkpoints, escalation rules, policy constraints and pause or resume control over a running agent. Another 406 document partial coverage, and only 66 document none.

Every vendor in the index is assessed against the same 14 point taxonomy from public documentation, and no vendor pays for placement. Counts on this page were measured across all 955 public vendors on October 7, 2026.

How the 955 vendors split

Full coverage483 vendors, 50.6%
Partial coverage406 vendors, 42.5%
No public evidence66 vendors, 6.9%

No public evidence means the reviewed sources did not document the capability. On this index that is a statement about the evidence, not proof that the capability is absent. See methodology.

What counts as full coverage

Full coverage means oversight is a configurable control the vendor ships rather than a property of one workflow: the buyer can decide which actions need sign off, who signs off, and what happens when nobody does. Partial commonly means a human handoff exists in one place, usually escalation to a live agent, or a boundary that is stated rather than enforced. A gate the buyer already owns, such as a draft pull request awaiting review in their own repository, is a delivery convention and not the vendor's mechanism.

How to read these numbers

This is one of the most evenly split axes in the taxonomy, and the split itself is the finding. Almost every vendor documents something, and only about half document a control. That is what an immature governance market looks like: escalation is built because customers complain, while approval policy is built because a security review demands it, and only the second survives an audit. Voice at 25 percent and GTM at 38 percent sit lowest, with agent infrastructure close behind at 41 and data analyst agents at 42, which is exactly where autonomous action is least supervised, so if you are deploying agents with real write access, read this axis before the capability axes.

Leading platform for human oversight & guardrails in each use case

Picked mechanically: the highest total coverage vendor in each lane that documents full evidence on this axis, one vendor per row. Scores are out of 14.

  1. 1. Appian, for enterprise operations agents

    14 / 14

    Full enterprise operations agents ranking · Compare the whole lane on all 14 axes

  2. 2. FLOWX.AI, for multi-agent platforms

    14 / 14

    Full multi-agent platforms ranking · Compare the whole lane on all 14 axes

  3. 3. Gumloop, for GTM and revenue agents

    14 / 14

    Full GTM and revenue agents ranking · Compare the whole lane on all 14 axes

  4. 4. Mastra, for agent infrastructure platforms

    14 / 14

    Full agent infrastructure platforms ranking · Compare the whole lane on all 14 axes

  5. 5. ServiceNow, for customer support agents

    14 / 14

    Full customer support agents ranking · Compare the whole lane on all 14 axes

  6. 6. UiPath, for agent builders

    14 / 14

    Full agent builders ranking · Compare the whole lane on all 14 axes

  7. 7. GitHub Copilot, for coding agents

    13.5 / 14

    Full coding agents ranking · Compare the whole lane on all 14 axes

  8. 8. Dataiku, for data analyst agents

    13 / 14

    Full data analyst agents ranking · Compare the whole lane on all 14 axes

  9. 9. Komodor, for SRE and DevOps agents

    13 / 14

    Full SRE and DevOps agents ranking · Compare the whole lane on all 14 axes

  10. 10. Adopt AI, for browser and computer-use agents

    12.5 / 14

    Full browser and computer-use agents ranking · Compare the whole lane on all 14 axes

  11. 11. ElevenAgents, for voice agents

    12.5 / 14

    Full voice agents ranking · Compare the whole lane on all 14 axes

  12. 12. Torq, for security and SOC agents

    12.5 / 14

    Full security and SOC agents ranking · Compare the whole lane on all 14 axes

  13. 13. Cohere Health, for healthcare agents

    10.5 / 14

    Full healthcare agents ranking · Compare the whole lane on all 14 axes

Documented coverage by use case

Share of each lane documenting full coverage on this axis. Vendors that sit in two lanes count in both, the same rule the rankings and matrices use.

Use case Full coverage Share
multi-agent platforms 34 of 52 65%
SRE and DevOps agents 23 of 37 62%
enterprise operations agents 184 of 301 61%
browser and computer-use agents 25 of 42 60%
agent builders 70 of 118 59%
coding agents 41 of 70 59%
customer support agents 63 of 115 55%
security and SOC agents 40 of 85 47%
healthcare agents 30 of 69 43%
data analyst agents 25 of 59 42%
agent infrastructure platforms 76 of 186 41%
GTM and revenue agents 52 of 138 38%
voice agents 21 of 83 25%

Recent verified changes from the vendors named above

Capability coverage is not a static picture. These are the most recent sourced change log entries for the platforms listed above, newest first, one per vendor. Scores on this page update as entries like these are verified.

  1. Gumloop human approval / guardrails

    Medium impact

    App Rules in Gumloop 13.2.0 now run on a decisions model. Admins describe in plain language what a tool call must not do, and the model judges each call against that rule to block it, tag it, or hold it for human approval. The same release lets users approve a background subagent's requests from the main conversation in chat, Slack, iMessage or email.

    October 28, 2026 · Partially Verified · All Gumloop changes

  2. Dataiku human approval / guardrails

    High impact

    DSS 15.1.0 adds a Plan mode to Cobuild, Dataiku's natural language build agent, which proposes a set of high level tasks that users review and refine before the build starts. Once the plan is approved, Cobuild switches to Build mode and tracks progress against it. The same release adds an activity monitoring screen for deployed agents and a Secrets Manager for credentials used in connections and code.

    October 8, 2026 · Verified · All Dataiku changes

  3. GitHub Copilot human approval / guardrails

    High impact

    Local sandboxing is now generally available in GitHub Copilot CLI, the Copilot app and VS Code sessions using Agent Host. Commands and tools the agent runs, including local MCP and language servers where supported, are restricted in file, network and credential access by developer or organization policy, enforced through Microsoft eXecution Container on Windows, macOS and Linux.

    October 7, 2026 · Verified · All GitHub Copilot changes

  4. ElevenAgents human approval / guardrails

    Medium impact

    Merge proposals are now in alpha in ElevenAgents. Teams can open a proposal for an agent branch, review the configuration changes and test results, leave comments or approval decisions, and merge approved changes into a target branch, with matching clients in the JavaScript and Python SDKs v2.71.0.

    October 5, 2026 · Verified · All ElevenAgents changes

  5. UiPath human approval / guardrails

    High impact

    Administrators can now bring their own safety vendor into UiPath's AI Trust Layer, in preview, so agent guardrails run under the customer's own vendor agreement and region. Azure AI Language, Azure Content Safety and Noma are supported, checks can be enforced across the organization, and each evaluation is recorded in the agent trace.

    October 1, 2026 · Partially Verified · All UiPath changes

Full change log · updated weekly across the whole index

Questions buyers ask

Is escalation to a human the same as an approval step?

No, and the difference matters. Escalation hands a conversation over after the fact. An approval step blocks an action until a person authorizes it. Many partial scores on this axis document the first and not the second.

Which lanes have the strongest oversight coverage?

SRE and DevOps at 72 percent, then multi agent platforms and customer support, both at 67 percent. All three are places where an unsupervised agent can do visible damage quickly.

Does a low score mean the platform is unsafe?

It means the public evidence is thin. Unknown on this index means insufficient documentation, not proven absence, and it is a reasonable thing to put directly to the vendor in a security review.

The other 13 axes

No single axis decides a shortlist. Buyers who care about this one usually check knowledge grounding & rag and security, identity & governance next, or open the full taxonomy to see how the 14 axes fit together.

Contact us

Found a vendor we missed? Have feedback on the index? We'd love to hear from you.