Agent Framework
Build agents that pursue one goal under declared rules. Each choice an agent makes is taken from a legal set its rules allow, every action it takes passes a policy, and every step is recorded so you can check it later.
Try it now
Talk to the sales assistant, a digital worker built on the framework, and watch what it decides on every turn. Nothing to install.
Get started
The Agent Framework is available. You build agents in the console or from an agent folder with the agent command-line tools, test them on sample data, and deploy them as an application; the user guide walks through each step. Console access is granted on request: email hello@helixor.ai with the subject "Agent Framework access". Capabilities that are not built yet are marked Planned where they appear.
What you build#
An agent has a goal, the state it keeps between steps, decisions made within rules, and an audit trail. Several agents deploy together as one application, versioned and rolled back as a unit.
| Part | What it is |
|---|---|
| Instructions and skills | The agent's job, voice and boundaries, and the reusable procedures it consults. These guide its behavior; they do not grant it any action. |
| Tools | The declared operations the agent may call. A governed tool runs only when the policy allows it, and can be held for a person to approve. |
| Decisions within rules | A playbook decides which moves are legal; a trained decision head chooses among them; it can never pick a move the rules forbid. |
| The record | One durable row per step: what was read, the legal moves, the chosen move and why, and the version of every component involved. |
Build an agent#
Create your first agent
Start in the browser or from an agent folder with the refund-helper starter and sample data.
User guideSkills and tools
Give an agent procedures to consult and declared operations to call.
User guideYAML reference
Every key in agent.yaml, application.yaml and the rules file.
Rules and approvals
Decide which tool calls may run and which wait for a person.
User guideDeploy and versions
Deploy an application, compare revisions and roll back.
User guideTriggers and history
Run agents on a schedule, an event or a workflow, and read what each run did.
Understand how it works#
Build a digital worker
What a worker is, when to choose one instead of the Decision Runtime in your own code, and how each capability is licensed.
ConceptsHow agents work
The loop of goal, perception, strategy, action and outcome, with the mechanism behind each step.
Governed actionsGoverned execution
The action broker, per-task policy, limits, hash-chained receipts and the kill switch.
Worked exampleHow the sales assistant was built
Decision heads, a playbook-bounded move set, the reply gate and the evaluation behind a production worker.
Try itTalk to the sales assistant
Hold a live conversation and watch, turn by turn, what the platform decided.
How it relates to the other products#
An agent's decisions use the same idea as Decisions as a Service: a choice taken from declared rules with a record of why. When a step is a routing or rostering problem, Solver as a Service answers it with a verdict, a solution and the reason when there is none.