Governed agent execution on the Scrums.com platform. Configure agents, define policies, grant tool access, execute runs, and audit every action — through a single, controlled gateway.
The Scrums.com API is planned and not yet publicly available. Endpoints and behaviour are subject to change before release.
The AI Agent Gateway is the controlled execution environment for all AI agents on the Scrums.com platform. Every agent invocation on the platform flows through the gateway, which enforces policies, controls tool access, routes requests to the appropriate execution context, and records a complete audit trail.Agents are not raw API callers. They operate within a governed runtime: their capabilities are declared, their tool access is explicitly granted, their outputs are logged, and every action is auditable. The gateway is designed for production use in enterprise environments where agent behavior must be predictable, auditable, and constrained.Every active agent deployment is a Service Line with execution_model: agent_runtime. Usage is metered in runs and tokens. All run outputs are traceable back to the RUN-* record and the originating AGT-*.
An agent (AGT-*) is a configured execution unit with a defined purpose, model, system prompt, tool access list, and governing policy. Agents are not general-purpose. Each agent is scoped to a specific task domain and its capabilities are explicitly bounded by its configuration and policy.
Tools are the capabilities granted to an agent — the external systems it can interact with. Tools are declared on the gateway and granted to agents explicitly via policy. Available tool types:
A run (RUN-*) is a single agent invocation. Runs have structured inputs, outputs, tool call logs, token usage, duration, and an outcome status. Every run is immutable after completion and is recorded in the Audit Logs.
Policies can require human approval before certain agent actions. When an agent reaches an approval gate, the run is paused with status: "awaiting_approval". A human reviews the pending action and approves or rejects it via the API. The run then resumes or terminates.
{ "workspace_id": "WS-26-000021", "subscription_id": "SUB-26-001122", "name": "Delivery Summary Agent", "description": "Generates weekly delivery summaries from task and metric data.", "model": "claude-opus-4-6", "system_prompt": "You are a delivery reporting assistant. When invoked, you retrieve task completion data and metric summaries for the specified workspace and period, then produce a structured delivery summary report.", "policy_id": "POL-26-000004", "tools": ["scrums_tasks", "scrums_observability", "slack_notify"], "project_tags": []}
{ "data": [ { "id": "AAUDIT-26-000041", "agent_id": "AGT-26-000012", "run_id": "RUN-26-001801", "event_type": "policy_violation", "detail": "Agent attempted to call tool 'jira' which is not in the permitted_tools list for policy POL-26-000004.", "action_taken": "tool_call_blocked", "occurred_at": "2026-04-12T14:22:00Z" } ]}
Policy before agent, always. Never create an agent without a policy. An agent without a policy cannot be activated. Design the access model before writing the system prompt.
Start with narrow tool lists and expand. Granting an agent access to every tool available is not a starting point — it is a governance failure. Start with the minimum set of tools the agent needs for its defined purpose.
Review the tool call log, not just the output. The output tells you what the agent produced. The tool call log tells you how it got there and whether it stayed within expected behavior. Both matter for production agents.
Use require_approval_for on destructive actions. Any tool call that modifies data in an external system — creating Jira issues, updating tasks, executing code — should be gated behind human approval until the agent has demonstrated reliable behavior in your context.
Audit policy violations weekly. A pattern of violations indicates either a misconfigured agent (system prompt is asking for capabilities the policy doesn’t grant) or a misconfigured policy (the agent legitimately needs a tool that isn’t permitted). Both need fixing.
