A suite agent works from the record you already trust it with
Lattice states it plainly: an AI Agent grounded in your company's performance and talent data, with skills for summarization, coaching, analysis, note-taking and composition (lattice.com). Workday's agents run on Agent Harness, which hands each one the person's role, reporting line, permissions and the policies that apply, taken from the system that runs HR (workday.com). Oracle's twelve use cases describe the same arrangement: agents reading employee records in HCM, finance data in ERP and corporate documentation, with people assessing the suggestions and free to overrule them (oracle.com).
Inside the record, that is the right design. A review summary drawn from the reviews, goals, feedback and 1:1 notes the suite already stores is work an outside agent cannot do as cheaply.
Headless access, read in the direction it was written
On 13 October 2026 Workday opened its HR and finance data to agents built with other tools and models. Every request carries Workday's rules with it: who is asking, what they may see, which approvals apply, how the action is recorded. Agent-Ready Tools are Workday's MCP tools, so any agent that speaks MCP can read Workday data and act through them (newsroom.workday.com). The release is clear about direction. Outside agents reach Workday, under Workday's permissions, while Workday's own agents run on Agent Harness. Nothing in it has a Workday agent creating a Google Workspace mailbox or revoking a session in Okta.
"Headless can't mean lawless," Workday's chief technology officer says in the same release. Agreed. The question for a buyer is narrower: which agent may act in the systems where the work finishes?
The other half of HR work sits in systems the suite does not own
A start date needs a mailbox, groups and aliases, Slack channels and a first-week project: four systems on day one, and the HR suite is none of them. A departure needs Okta, the GitHub organization seat, ownership of the person's Drive files, the license. A policy question needs the inbox, the policy library, and the channel where an escalation gets picked up.
Each step is small, and each carries a permission a suite agent cannot grant.
An agent you host can grant it. Connectors reach the inbox, the identity provider, the drive and the repository, with 3,000+ apps in a click, plus MCP, OpenAPI, GraphQL, and raw HTTP where a system needs a narrower path. The policy desk page shows the pattern on an inbox where the agent drafts and a person sends.
Approval has to cross the boundary with the work
Workday's control is the one HR already trusts, and the release is specific about it: approvals and audit trails ride with every request. Early access in early 2027 adds Grounded Reasoning, which checks a request against company policy and tenant data before an agent acts, with the example of a manager asking to move someone into a role they have no authority to fill. Lattice puts the Agent's output in front of managers and HR to act on. Oracle's page puts it as people assessing the output and overruling a recommendation.
Kortix takes a different route to the same principle. Approval gates you set. Off until you set them.
The difference shows up when a write lands outside the HR suite. Approval holds the call; the agent's turn pauses and resumes. The grant, the rule and the argument that triggered it sit in the project, so the record of an approval sits next to the change it allowed. The governance page covers those files.
Where it runs decides most of the residency question
Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Workday runs its agents on its own platform, and Lattice's layer is embedded in Lattice. A company with a works council or a data residency rule asks this before it asks about features, and the answer is a deployment it controls.
When the suite's own agent is the right pick
Three cases where the suite's agent fits. If nearly all HR work happens inside Workday records and you want Workday's rules applied to it, its harness is the shortest path, and the headless tools make it reachable from your own build. If performance, growth and engagement data live in Lattice, its Agent reads context an outside agent would have to reconstruct. If the HCM and the ERP are Oracle, the twelve use cases map onto systems you already run.
Then the work crosses a boundary. A new hire needs a mailbox and a Slack invite. A departure needs the identity provider. A policy answer needs the inbox. Once more than one of those is true, the agent has to leave the suite, and the question becomes what it carries with it.
What stays a company decision
The policy text, the bar for escalating a question, the rubric behind a screening score and the list of calls that need approval are yours. No vendor supplies them. What changes between the options is where the mechanism lives: in a product you rent, or in files you can open. Open source. Read it, fork it, audit it.