01
Design
Create a workspace and define skills, contracts, transitions, commands, and constants.
AI Agent Management System
Manage AI agents as business processes
Hollon AMS brings AI agent design and orchestration into one workspace with skill graphs, contracts, versions, collaboration, and analytics. External AI agents retrieve the process and report execution through Remote MCP.
Markdown + YAMLGitHub and GitLabRemote MCP
Request handling
Primary process
OverviewGraphCommandsConstants
Accept request
intake-request
Build plan
build-plan
Execute
execute-task
2 inputs1 output
Review
review-result
Contracts are compatible
Why AMS
Turn scattered prompts into a manageable process
When instructions, transitions, and data live in separate files, teams discover incompatible steps too late and lose sight of the process as a whole. AMS turns that environment into an explicit, versioned system.
One AMS workspace represents one agentic business process — from skill definitions to run history and business metrics.
Visible structure
Skills, allowed transitions, selection reasons, and missing targets are all shown on one graph.
Explicit contracts
Required and optional inputs and outputs are defined for every step.
Shared rules
Commands and non-secret constants are reused across skills without duplication.
Observable execution
Related runs are grouped into process runs, providing the foundation for analytics and KPIs.
How it works
Design, run, measure, improve
Process definition and execution stay separate: AMS manages structure and history, while the model runs in an external agent environment.
02
Connect
Choose a version and connect an external AI agent through Remote MCP with the appropriate scopes.
03
Run
The agent retrieves requested materials, executes skills in its own environment, and records the results.
04
Measure and improve
Compare versions, identify bottlenecks, and link skill outputs to business KPIs.
A controlled lab for continuous experimentation
Features
Every process component in one workspace
From individual skill structure to team permissions and Git synchronization for the primary version.
Design
Define the process as a system
Interactive skill and transition graph
Markdown, metadata, and file editor
Input and output contracts
Shared commands and non-secret constants
Control
Collaborate without losing history
Editable versions and comparisons
Viewer, Editor, and Administrator roles
Comments and activity history
30-day recycle bin with restore
Integrations
Connect and move your process
Remote MCP over Streamable HTTP
OAuth PKCE and personal API tokens
Workspace import, export, and duplication
GitHub or GitLab for the primary version
Versions and collaboration
Evolve the process without losing reproducibility
Teams can edit multiple versions in parallel. Each user independently selects a working version, while the primary version is used by default and synchronized with Git.
A new version copies the process definition, element layout, and KPI mappings — without run history or comments.
A process run keeps the selected version and definition state captured when it started.
An archived version remains available for history and comparison, but cannot accept changes or new runs.
Process versions
Request handling
v3 · Working
Primary versionUpdated
today, 2:32 PM
History
128 runs
OpenCompare
v4 · New fallback
In progressUpdated
today, 12:08 PM
History
7 runs
OpenCompare
v2 · Before optimization
ArchivedUpdated
July 28, 6:41 PM
History
94 runs
OpenRestore
Process is pinned
Changes to the primary version do not redirect a process that is already running.
Remote MCP and an open format
Connect external agents without locking the process to one environment
AMS exposes process definitions and history over Streamable HTTP. Models and scripts run in the external agent under its own security policies.
External AI agent
Codex, an IDE, or a headless client
requestRemote MCP
AMS server
Roles, scopes, and revisions
PostgreSQL and storage
GitHub or GitLab
Two access modes
OAuth Authorization Code with PKCE or a personal API token.
Restricted scopes
A connection can never receive more access than the user’s role allows.
Only on request
The client receives the available materials it requests, not the entire workspace automatically.
Portable structure
Markdown + YAML.ams/
└─AMS.md
└─skills/
└─skill-name/SKILL.md
└─commands/
└─constants.json
Import an `.ams.tar` or `.zip`, export the entire workspace, and keep the primary version in Git. Scripts remain process materials — AMS does not execute them.
Analytics and KPIs
Measure the process, not isolated calls
Related runs are grouped into process runs, so statuses, transitions, errors, duration, bottlenecks, and business KPIs can be analyzed by version and initiator.
Process overview
Customer support
Last 30 daysCompare versions
Process runs
1,248
+12% this period
Success rate
94.8%
1,183 completed
p50 / p95
42s / 2.1m
across all versions
Bottleneck
review
38% of total time
Skill performance
Primary version · v3| Skill | Runs | Success | p50 |
|---|---|---|---|
| classify-request | 1,248 | 99% | 3.2s |
| resolve-request | 1,121 | 95% | 27.4s |
| review-result | 1,066 | 91% | 48.8s |
Business KPI
First-contact resolution
78.4%
Source: the `resolution_status` output of the `resolve-request` skill in v3.
If a skill or its output is missing from the selected version, AMS clearly marks the KPI as unavailable for that version.
Security
Clear responsibility boundaries from day one
AMS manages process definitions, access, and history without replacing the agent runtime or a secrets management system.
Roles and scopes
Team membership and roles are checked for both the web interface and MCP. Connection scopes can only narrow permissions.
Secrets stay outside
Tokens and keys must not be stored in constants.json, instructions, Git, or exported archives.
Execution stays separate
AMS stores and delivers scripts as process materials, while only the external agent runs them in its own environment.
Transparent telemetry
AMS does not receive the model’s hidden reasoning. Extended connection data is retained for no more than 30 days.
Who it’s for
For teams bringing agents into real-world workflows
AMS gives business owners, engineers, and external agents a shared language — from process design to measurable outcomes.
Example processes
Feature delivery: analyze → implement → verify
Support: classify → resolve → escalate
Research with multiple continuation paths
Content: generate → review → approve
AI engineering teams
Align on process architecture, work across versions, and collaborate on individual skills.
Technical leaders
Control contracts, access, reproducibility, and error-handling policies.
Developers and integrators
Connect agents through MCP, synchronize with Git, and move processes between environments.
AI product owners
Connect technical runs to business KPIs and see where the process needs improvement.
FAQ
Questions for technical evaluation
Looking for another detail? Request a demo and we’ll walk through the product using your scenario.
Hollon AMS is an AI agent management system for versioned business processes. It brings skill graphs, data contracts, collaboration, Remote MCP, and outcome analytics into one workspace.
One workspace represents one agentic business process. It brings together shared rules, skills, data contracts, transitions, versions, permissions, and analytics.
No. AMS manages process definitions and history. Models and scripts run through an external AI agent in its own environment and under its own security policies.
Through Remote MCP over Streamable HTTP. Interactive clients can use OAuth Authorization Code with PKCE, while headless scenarios can use personal API tokens. User roles and connection scopes restrict access.
A workspace can have several versions being edited in parallel. One is designated as primary and used by default; only that version synchronizes with GitHub or GitLab. A process already in progress stays on its selected version and captured definition state.
Yes. The portable part of a process uses an open Markdown + YAML structure. A workspace can be imported from TAR/ZIP, exported, duplicated, and stored in Git.
Telemetry is configured per connection. Extended mode may include arguments, MCP operation results, and trace events; this data is retained for no more than 30 days. AMS does not receive the model’s hidden reasoning.
Build a manageable process for your AI agent
Open Hollon AMS, select a working version, and connect the process definition to real execution through MCP.