For model risk and second line
Can we govern it, and show that we did?
Govern it, and show that you did.
Each completed agent run records which model answered, each decision is kept, and each run ends with a verdict written to a hash-chained ledger. Below, the questions a review asks of Promenaut, with what answers each on the platform.
The governance pack: Governing the Promenaut Workforce and Model Risk Evidence on Promenaut, sent on request.
- Model and providerrecorded for each completed agent run
- Approvalwho decided, item by item
- Roster verdictthe agents that ran, against the roster
- Ledgerhash-chained, so a later edit shows
The questions, in the order they come.
- An agent's declared boundary, drawnEach agent's declared boundary is shown as one card with the agent at its centre: the workflows and platform teams that admit it, the agents declared to call on it, the service teams it may call, its declared effects and invocation surfaces, the knowledge it reads, and its preferred model and clearance, with anything the agent does not declare shown as not declared.
- Declared I/O contractsAn agent declares what it consumes and produces.
- Versioned, diffable agent configurationAn agent configuration is a versioned object with its own lifecycle: versions can be listed, compared against each other, and published deliberately rather than taking effect on save.
- Approval gate on agent workWork can be held for a human decision, approved or rejected individually or in bulk, with the decision history kept and retrievable per workflow.
- Run quality never suppresses governanceWhether a piece of work needs human review is decided by declared configuration alone.
- A trust gate between an agent and its changesAn agent can pass a change it has found through a trust gate instead of making it.
- Masked content is delivered; a block says why without saying whatWhen an Amazon Bedrock guardrail checking a whole call masks sensitive content rather than blocking it, the masked response is delivered and the work continues; when it blocks, the call fails with a reason that names only the kinds of policy and the categories that fired, never the matched text, and the block is not retried or passed to the fallback model.
- Output egress wallWhere a regulated workflow is provisioned to read held or restricted content, so that its conclusion is actually correct, a second wall applies to what it produces: the deliverable, the live event stream and the audit record must not reproduce that content or reveal that a hold exists.
- Four-eyes on sensitive knowledge changesA domain can be marked as requiring approval.
- Run-end roster verdictWhen a workflow run ends, whether it completed, failed, was stopped, was cancelled or was ended by a containment rule, the platform records a content-free pass, fail or no-baseline verdict in the audit ledger on whether the agents that ran matched the roster the run launched with, and a run in which an agent called a service team its module did not declare also fails.
- Agent call monitoringEach agent that ran inside a workflow is classified against the roster that run launched with: expected, adaptive (added by a rule that fired in the run), platform-enabled (a platform agent the workflow switches on, or one the platform runs on workflows by default, and the organisation has not withdrawn), delegated (run by a service team an approved agent is declared to call), unexpected, or no baseline, with a drill from workflow, each workflow on its own page, to run, each run on its own page, to agent.
- Declared versus actual conformanceAfter each agent run completes, a validator compares the agent's declared knowledge-graph write effect and its required knowledge outputs with a fresh query of the rows the run actually wrote, and records a content-free conformant, non-conformant or no-baseline verdict in the audit ledger; an agent that declares neither effects nor a required knowledge output is validated as no baseline.
- Tamper-evident risk ledgerModel risk events (suggestions offered and decided, the roster each run launched with, adaptive rules that fired, run-end roster verdicts, completion-validation verdicts, and changes to platform agent approval, to the rating prompt and to a workflow's monitoring switch) are written to the append-only, hash-chained audit ledger.
- Administrative audit trailSecurity-relevant administrative changes are written to a hash-chained audit log with the actor, address and session recorded, covering users, roles, teams, access policies, single sign-on, security mode, cost centres, each agent's knowledge envelope and service identity, the definitions of domain, platform and service teams, and workflow configuration settings.
- Agent lifecycle audit logEvery lifecycle transition of an agent is recorded with its previous and new state, the reason given and who made it, and the log is shown on the agent and as a fleet activity feed.
- User reliance trackingCo-authoring surfaces record content-free ledger events when a suggestion is offered and when a person accepts, modifies or rejects it; an unanswered suggestion counts as ignored.
- Reliance evidence exportThe reliance figures can be downloaded as CSV or Excel, as the overall breakdown or as one workflow or person drill, so a reviewer can take the evidence out of the console.
- Workflow-run ratingsWhen a workflow run ends, the person who ran it can rate it once: a usefulness score from 1 to 5 for a delivered run, or an outcome of failed or abandoned with no score, with an optional comment of up to 2,000 characters.
- A module states its own governance positionA module declares its purpose, intended use, data classes, human oversight controls, limitations and validation position in its manifest, and that block feeds the governance sections of the module's delivery pack.
What the platform does for you, layer by layer.
Requirements and delivery
- Delivery Design
Home and conversation
- The chat says it is an AI
- Run history across workflows
Deliverables
- Dashboards read from knowledge graphs as the viewer
- Ready-made dashboards a module offers each organisation
Actions out
- An agent reads from an approved outside MCP server through one checked, recorded call
Model gateways
- Every model call through the client's own AI gateway
Kinds and roles
- Deterministic and hybrid agents
- Intent is worked out from how an agent is built
- Try an agent on different models and prompts, and compare the results
- Score a workflow team's outputs with the platform evaluator
Models and routing
- Agents stay bound to models that exist
- Move agents from one model to another in one step
- Compare models side by side on real agent work
Output checks and repair
- A quality controller is built and not yet wired in
- A model-output validator is built and not yet wired in
Guardrails and model limits
- Masked content is delivered; a block says why without saying what
Documents and media
- An unreadable upload stops the run instead of passing silently
Registry and versions
- Versioned agent registry
- Versioned, diffable agent configuration
Fleet and control plane
- Per-agent control plane
- Readiness-first agent roster
- Fleet lifecycle attention queue
Teams and workflows
- Domain, platform and service teams
- One registry and one profile for every kind of team
- One registry and one profile for each workflow
- Each organisation sets its own values for a Workflow, and a run keeps its copy
- Workflow Lab: the anatomy of a workflow
Conversation to work
- Conversation resolves to declared capabilities
Execution
- A workflow cannot be marked complete with agents still running
Run control
- A running workflow can be paused, stopped or cancelled
- Workflow replay
- Stopping, pausing and resuming work that is already running
- A run that asked a person waits in a status of its own and resumes on the answer
People in the loop
- Human in the loop
- A module can ask the person who launched the run, and the run waits
- The platform asks like a specialist
Graph and search
- How far the knowledge graph can be trusted
Domains and scopes
- Sealed compartments and legal holds
- Firm-wide reference knowledge
- Knowledge graphs can be created, loaded, checked for health and archived from one page
Access decisions
- A read names who and why
- Per-agent knowledge exceptions
Governance controls
- What a firm can switch off, and what it cannot
- Classification an agent cannot choose
- Four-eyes on sensitive knowledge changes
- Output egress wall
- Break-glass access
- Knowledge access auditing
- Wrong data can be taken out of a knowledge graph and its earlier values put back
Ingestion and provenance
- Extracted knowledge says where it came from
- Source artefacts become typed knowledge
Promotion and submissions
- Each domain chooses how versions accumulate
- Governed promotion to a wider audience
Identity and access
- Service identities are reviewed on a cadence
- Administrators act only within their enterprise and authority
Approvals, decisions and containment
- Approval gate on agent work
- Proposed change inbox
- Dialled autonomy
- Run quality never suppresses governance
- Decision explorer
- Agent handbrake
- Suspend an agent everywhere
- One protocol for decisions a person must make
- No outside MCP server reaches an agent until it is approved by a second person, or by a super admin who registered it
- Platform staff can block an outside MCP server for every organisation at once
- An agent runs an approved public skill in a sandbox with no outside network, and its uses are recorded
- A trust gate between an agent and its changes
- Containment rules drain, quarantine, pause or block matching work
- One governed door for stopping and containing work
- A held run waits for the decision that holds it
- A second person approves a proposed directive
- Brakes on automatic control
- Containing work by what it touches
- Runaway agents are suspended automatically
- An agent that keeps failing on its own configuration is contained automatically
- One place to govern MCP, in both directions
- One checked, recorded door for the tools an agent offers a model
- A control directive can slow matching work down without stopping it
- A control directive can pause, drain or kill one agent or one team inside a running workflow
- One Approvals screen for what the control engine is waiting on
- An approval gate holds matching runs or calls until a person decides
- Automatic step-up of containment for a runaway agent
Model risk
- Contract truthfulness monitoring
- Model risk oversight console
- Agent call monitoring
- Run-end roster verdict
- Declared versus actual conformance
- User reliance tracking
- Workflow-run ratings
- Tamper-evident risk ledger
- Model inventory
- Model provenance and self-drift
- Agent run health and trends
- Model risk access audit
- Reliance evidence export
- Run evidence export
- An agent's declared boundary, drawn
- A diagram of one run for auditors
- Conversation oversight: replay a person's chat
- Whether each agent, team and run passed its model risk checks
- Pausing and stopping work is requested through control directives, not through model risk
- Policy rules that record, propose or apply a control directive
- A record of when each model was reviewed, by whom, and the outcome
- One cost figure for each run and agent, with its parts by kind
- One finding for failures that share a cause across the fleet
- Judging whether a run stayed within the intent it was launched with
Performance and review
- Per-agent and per-version performance
- Agents are reviewed on a cycle
- Reviewed by a manager, a peer, a person, or the platform
- Promote, maintain, or put on a plan
- An improvement plan with a target and a deadline
- The cycle starts itself
- Agent grading
- Agents evaluate other agents' work as it runs
Audit trails
- Administrative audit trail
- Agent lifecycle audit log
- Seam calls leave an activity record
- Old capability audit records are archived before they leave the database
- Audit ledger anchored outside the database
Credits and budgets
- What a run has spent, while it runs
Agent contracts
- Declared side effects
Modules and manifests
- A module states its own governance position
- What is inside each installed module
Agent building blocks
- Shared weighted scoring and conformance scorer base
- A conversational answerer is a provider on a shared loop
Authoring and testing
- A design is checked for completeness, reuse and how each agent will be proven
- A module design pack for review
Deploying environments
- A client deploy is a reviewed change
- Tearing down an environment needs approval
Operations and observability
- An operations console for running work and the agent fleet
Weighing us against the tools you already know? See how we compare, with sources →
How the second line’s questions are answered for agents, run by run. Model risk for AI agents →
The full answers, written for you.
Governing the Promenaut WorkforceAbout 20 minutes
Model Risk Evidence on PromenautAbout 30 minutes
Bring your model risk team.
Walk a live run from launch to verdict: what ran, what was approved, and the evidence the platform records.
Promenaut
