Define agent roles
Turn scenarios into runnable agents with role, capability, and trigger.
Agent Runtime · Build · Run · Govern · Observe
For enterprises ·
multi-agent collaboration · business automation
build · run · govern · observe
- 01
BUILD
role · capability · trigger
- 02
RUN
runtime · tools · commands
- 03
GOVERN
rbac · audit · breaker
- 04
OBSERVE
session → tool → token → cost
Four capabilities
Define the roles on the runtime first, then let those agents work together. Who may touch what is set in advance, and every step can be looked up afterward.
Make agents get work done
Execute on the runtime: tools, commands, and data changes.
Permissions, audit, breaker
Decide who can do what; trace incidents; stop out-of-bounds runs. High-risk actions can require human confirmation.
See traces and cost
session → tool → token → cost on one chain.
Build · Define an agent role
role + capability + trigger to turn a scenario into a runnable agent.
Run · Trigger a job
Event or schedule kicks off execution on the runtime.
Govern · High-risk confirm
Permissions, audit, breaker; confirm only high-risk actions.
Observe · Trace
session → tool → token → cost on one chain.
Auto-play · hover to pause · click for full size
personas
Who it's for
This runtime is not tied to one industry. The cards below are built-in role templates, and a role can join a collaboration. An industry that is not listed starts the same way: split the work into roles.
- Capacity & scheduling analyst
- Delivery performance analyst
- Material readiness analyst
- Metrics / systems
- Schedule attainment · OEE · Downtime hours · OTIF · Kit completeness
- Gates
- Read-only: no schedule changes, no purchase orders
- Supplier compliance checker
- Supplier on-time rate analyst
- Supplier risk grader
- Metrics / systems
- Certificate type & validity · On-time delivery rate · Incoming defect rate
- Gates
- An expired certificate means no admission; writing back a risk grade needs human approval
- Reconciliation break analyst
- Policy clause checker
- Regulatory filing reviewer
- Metrics / systems
- Reconciliation tolerance · Break direction · Clause no. + version + effective date
- Gates
- Above tolerance only yields "needs review", never a write-off; no clause means "not covered by policy"
- Change risk reviewer
- Incident root cause analyst
- SLO burn-down analyst
- Metrics / systems
- Change class · Approver count · Change window · Error budget
- Gates
- High-risk changes need two approvers inside the change window; "not excluded" is never written as "root cause"
- Data query assistant
- Knowledge Q&A specialist
- SRE on-call
- Office document assistant
- Metrics / systems
- 8 cross-industry roles · No industry system or metric binding
- Gates
- Interactive queries require human confirmation; production changes like restarts go through approval
- What it covers
- Role and capability development · mapping business and digital assets
- How it gets done
- Map the processes and the materials they need, then build the roles, capabilities, and approvals so those roles finish the work together.
Role splits, metric definitions, and thresholds are our own design examples, shown to illustrate capability boundaries—not a reproduction of any vendor case study.
scenarios
Scenario demos
Three recordings, three shapes: multi-role collaboration (parallel evidence gathering into a single context), orchestration with governance (sequential relay plus a human signature at the end), and single-role with an event gate (an SRE on-call triaging an alert, where changes like restarts stop for human approval). click a shot on the right to jump.
Quarterly subscription revenue review
Two workstreams gather evidence in parallel and converge into one context. The supervisor accepts each deliverable, and a rework also invalidates downstream work.
- Two parallel workstreams
- Cross-wave gating
- Upstream artifacts mounted
- Per-deliverable acceptance
- Rework invalidates downstream
- Pre-authorized approval trail
Shots
Click a shot to jump; the current shot highlights during playback.
How it works
Pick a step to see the matching console screens.
Create role
Name, persona, and runtime binding.
Capabilities
Attach tools, skills, and MCP.
Permission policy
allow / ask / deny; high-risk needs confirm.
Auto-play · hover to pause · click for full size
govern
Govern
Day to day this is permissions and records. When something goes wrong, the breaker stops the run. Human confirmation is only for high-risk actions.
-
Audit trail
Who triggered, approved, and spent—traceable, with chain verify when needed.
-
Breaker
When something goes wrong, the run can be stopped.
-
Permissions & confirm
Role-level allow / deny; high-risk actions can require human confirm—risk-triggered, not every step.
Audit trail
Who triggered, approved, and spent—traceable.
Breaker
When something goes wrong, the run can be stopped.
High-risk confirm
Only actions marked high-risk—avoid approval fatigue.
Waiting for approval
Execution pauses; tickets show up in Approvals.
Auto-play · hover to pause · click for full size
Integrations
Agents built here and external harnesses share the same observe and govern plane.
OpenCode
harnessLocal hooks report actions into the same observe and govern loop
Claude Code
harnessExternal execution actions join the same control plane
Codex / LangGraph
frameworkHomegrown or framework agents plug in by reporting actions
MCP / hooks
protocolProtocol adapters; execution stays in the customer's environment
Plans
Trial gets observe and basic permissions running; Enterprise completes governance.
Trial
Observe + basic permissions
- ·traces / cost panel
- ·Basic roles & permissions
- ·Console & self-host trial
- ·Self-confirm (HITL) sample
Enterprise
Full governance
- · Audit chain & verify
- · Breaker & cross-session govern
- · High-risk confirm & trail
- · Multi-tenant + SSO (roadmap)
Walk one full trace
Register in the console, configure a role, and walk one full trace.
FAQ
An enterprise runtime for multi-agent collaboration. Roles are defined on the runtime, then those agents carry out the work together. Who may touch what is set in advance, and every step can be looked up afterward.