AI Workforce Management

Define What Your AI Agents Can Do.

Shepherd is the management plane for autonomous agents. Define workload profiles by role, enforce tool permissions, and audit every action — with zero ambiguity.

Join Waitlist

Before Shepherd vs. With Shepherd

Most agent runtimes ship with static tool allowlists and hope for the best. Shepherd scopes your agent's access by what the job actually requires.

Category	Without Shepherd	With Shepherd
Tool Access	Static allowlists — on or off	Role-scoped permissions by job function
Scope Definition	Manual configuration per agent	O*NET-derived workload profiles
Enforcement	LLM-based — agent 'promises' to comply	Deterministic — binary ALLOW or DENY
Audit Trail	Scattered logs, no central view	Every action logged with decision rationale
Boundaries	Hope the agent stays in scope	Filesystem fences, exfiltration blocking, approval gates

Built for Production

Deterministic enforcement built for speed, security, and transparency.

Agent Permissions

Define what each agent can do. Set allowed tools, blocked actions, and approval requirements. Permissions enforced automatically on every action.

O*NET Occupation Data

Thousands of task statements across 923 occupations from the Department of Labor, mapped to tool permissions. Pick a role, get a policy. No guesswork.

Real-Time Feed

Watch every tool call, every decision, every boundary check as it happens. Color-coded ALLOW and DENY with latency measured on every request.

Approval Gates

Sensitive actions like sending emails or modifying calendars require human confirmation in Slack before execution.

Channel Routing

Map Slack channels to agents with different permission profiles. #admin gets full access. #content gets read-and-write only.

Deterministic Enforcement

No LLMs checking LLMs. Binary ALLOW or DENY decisions from deterministic rules. Same input, same decision — no model drift, no exceptions.

Policy as Code

Shepherd defines agent permissions as structured policy — allowed tools, denied actions, approval gates, and boundary rules. Every tool call is evaluated deterministically against your agent's permitted scope. No inference, no guessing.

1package shepherd.profiles
2
3# Marketing Manager — Workload Profile
4# Source: O*NET 11-2021.00
5
6default allow := false
7
8# Allowed tools for this role
9allowed_tools := {
10    "read", "write", "web_search",
11    "email.draft", "calendar.read"
12}
13
14# Tools requiring human approval
15approval_required := {"email.send", "calendar.modify"}
16
17# Denied tools — outside workload scope  
18deny {
19    not allowed_tools[input.action]
20    not approval_required[input.action]
21}
22
23# Filesystem boundary
24deny {
25    startswith(input.target, "/etc")
26}
27
28# Exfiltration protection
29deny {
30    contains(input.target, "webhook.site")
31}

Example policy structure — enforcement uses the same allow/deny/approval rules in our deterministic engine.

Ready to manage your AI workforce?

Define workload profiles, enforce tool permissions, and audit every agent action — from a single dashboard.

Join Waitlist →