AI AGENT ASSURANCE
DriftGard tests agent behaviour against reviewed business expectations and flags regressions before release. Compare releases and repeat checks on a schedule, with evidence for your next decision.
Reviewed scenarios · Release comparisons · Scheduled checks
DriftGardREFUND AGENTDGAGENT ASSURANCE
scenarios passed
Illustrative refund workflow
GOVERNANCE ACROSS THE AGENT LIFECYCLE
From ownership and risk to testing, release decisions and execution evidence. Connect the capabilities your workflow needs.
Record business and technical owners, purpose, data sensitivity and permitted actions. Document your organisation’s risk assessment and rationale.
Explore ownership & releaseTest reviewed scenarios, compare releases and repeat trials. Bring the evidence to designated reviewers, with approval conditions and the selected policy and models recorded.
Explore Agent AssuranceApply policies, tool limits and human approval at connected action boundaries. Configure release clearance checks for deployment and live requests using the supplied release context.
Explore RuntimeFollow captured activity and configured alerts. Track AI costs and estimate business value from reported outcomes.
Explore visibilityConnect versioned policies, release decisions, approval conditions and clearance history with evaluation evidence. Share reports and structured records.
Explore evidenceChoose the capabilities your workflow needs; these are complementary, not mandatory steps. Testing evaluates behaviour; connected Runtime paths apply controls; monitoring and evidence support oversight.
AGENT ASSURANCE
Your agent changed. Did its behaviour improve? Test the scenarios that matter to your business, compare against a baseline, and record the decision to approve or reject a release.
One failed scenario. Three failed trials. The baseline agent proposes a $100 refund instead of requesting approval.
6 scenarios × 3 repetitions = 18 trials · 15 passed · 3 failed
OWNERSHIP & RELEASE GOVERNANCE
Make accountability part of every release. Know who owns the agent, what was reviewed and which policies and models its clearance covers.
Record owners, purpose, data sensitivity, permitted actions and the rationale for its risk classification.
Who is accountable?Test reviewed expectations, compare against a baseline and keep failures and unresolved results visible.
What was tested?Route evidence to designated reviewers. Record decisions, conditions, the Control Pack version and approved models.
Who approved what?Record clearance once conditions are satisfied. Bind application keys and configure deployment and live-action checks.
Is this release cleared?A passing test does not automatically approve a release. Clearance enforcement depends on configured boundaries and the supplied agent, release and environment context.
See the release workflowDRIFTGARD RUNTIME
Carry your business rules into execution with tool limits, policy checks and human approval. Runtime applies configured policies to connected model, tool and final-response paths.
DriftGard RuntimeA controlled execution environment
TRY A POLICY DECISION
Example policy: allow under $100, request approval from $100 to $500, block above $500.
The request is within the agent’s authority. The connected tool can proceed, with the decision recorded.
Next: execute the tool and record its outcomeYour environment. Your controls. Deploy compatible workloads in your infrastructure, with model and tool credentials held outside the agent workload.
Explore Runtime for your teamGOVERNANCE VIEWS
Connect agent activity, policy decisions and reported outcomes. Give each team a clear view of what needs attention.
Follow allowed actions, blocked requests and recorded human reviews.
SIX VIEWS. ONE PROJECT.
Investigate what was allowed, what was blocked and which rule made the difference.
Agent, tool, policy and review context.
Track reported spend and usage-based estimates. See costs by agent or model, including gaps in coverage.
29 model calls · 28 priced · 1 unpriced
Keep reported and estimated costs distinct.
Use recorded execution outcomes and your operating assumptions to estimate time saved and net value.
Value from successful work, with visible assumptions.
These charts illustrate a refund workflow using simulated usage and sandbox execution outcomes. Policy evaluations and model calls are counted separately; their totals need not match. Cost uses illustrative model rates. Value assumes 7 minutes saved per successful execution, AUD 60 per hour and AUD 0.05 per measured execution. Figures are examples, not customer results.
GOVERNANCE EVIDENCE
Bring the approved policy and models, release decision, reviewer and recorded outcome into view. Give technical teams context to investigate and stakeholders evidence to review.
BUILT AROUND YOUR WORKFLOWS
Choose where you need control and where you need visibility. Each connection has a clear role.
Test agents handling refunds, support tickets and business records. Set tool limits and require human approval for sensitive actions.
Permissions · Tool limits · ApprovalsReview captured activity across connected workspaces. Investigate policy findings by agent, tool and developer, with the surrounding context.
Activity · Policy findings · ContextCheck a proposed action before your application executes it. Your integration enforces the returned decision and reports the outcome.
LangChain: Python & Node.js.
CrewAI and Strands: Python.
Run a compatible workload in your environment. Runtime governs the connected model, tool and final-response paths, with credentials outside the agent workload.
You operate the deployment. Policies apply at the configured Runtime boundaries.
Collect recorded coding-agent activity for policy evaluation. Observation is retrospective: it does not block commands or hold them for approval.
Codex, Claude Code, Cursor, Kiro IDE and GitHub Copilot in VS Code. Capture coverage depends on the host and session format.
A FEW USEFUL ANSWERS
Bring one agent and one workflow.
We’ll help you map the rest.
Start with Agent Assurance to test behaviour and compare releases. Start with Runtime for governed access during execution. They work together: Assurance checks expectations; Runtime applies controls at connected action boundaries.
Start with an agent risk profile: accountable business and technical owners, purpose, audience, data sensitivity and permitted actions. The risk classification records your organisation’s assessment and rationale; it is not an automatic risk score for every action.
Test reviewed expectations with Agent Assurance, then bring the evidence, Control Pack version and approved models into a release review. Designated reviewers record approval, conditions or requested changes. Record clearance once conditions are satisfied, then configure deployment and live-action checks.
Approval records do not deploy an agent or prevent actions by themselves. Enforcement requires the configured boundaries, bound application keys and supplied release context. Keep your existing security, legal and risk review responsibilities.
People pay for insurance for peace of mind when outcomes are uncertain. DriftGard brings that mindset to AI agents: ongoing checks against business expectations, controls for sensitive actions and evidence for accountable decisions.
Agents in production are inherently unpredictable. A successful demo or a passing assessment cannot make an agent 100% reliable. DriftGard helps you identify changes in tested behaviour, require human approval where needed and review what your connected application reports happened. You can make decisions with continuing evidence rather than relying on a single sign-off.
“Agentic insurance” is our analogy for managing agent risk. DriftGard provides governance software rather than an insurance policy that pays claims.
LangSmith supports tracing, offline and online evaluations, regression comparisons and custom evaluators that can check business rules. Keep that workflow. DriftGard combines reviewed agent assessments with versioned Control Packs, human approval workflows and connected decision and reported-outcome evidence.
Start with one business workflow: does connecting its expectations, action limits, approvals and evidence simplify a process your team currently maintains separately? Add DriftGard where that closes a useful gap. Connecting platforms requires integration; DriftGard does not automatically import LangSmith tests or synchronise policies.
Microsoft Agent 365 provides a control plane for agent identity, observability, security, governance and lifecycle management, using services including Entra, Defender and Purview. DriftGard focuses on reviewed business expectations for a connected workflow, with agent assessments, versioned policies, approval records and evidence linking decisions to reported outcomes.
Keep your existing identity and security controls. Evaluate whether DriftGard closes a workflow-specific assessment or accountability gap across the agents and systems you connect. Agent 365 and Microsoft Foundry evaluations are distinct capabilities; neither is automatically connected to DriftGard. Agree the integration and enforcement boundaries before adopting another platform.
Amazon Bedrock AgentCore offers runtime infrastructure, evaluations and observability. Its Policy capability evaluates requests through AgentCore Gateway before tool access, including rules based on identity and tool parameters. These capabilities overlap with parts of DriftGard.
DriftGard brings reviewed business assessments, Control Packs, human review workflows and connected governance evidence together for the agents and systems you integrate. If your AWS setup already meets those needs, assess the remaining gap first. DG SDK/API integrations require your application to enforce decisions at configured boundaries; using AgentCore does not automatically connect those boundaries or synchronise its policies with DriftGard.
Explore AgentCore Policy ↗ · Explore AgentCore Evaluations ↗
No. Agents in production are inherently unpredictable. An agent that passed yesterday can break a rule tomorrow. Models, prompts, tools, data and business processes change, and production introduces situations your tests may not cover. A passing assessment shows that the reviewed scenarios met their assertions on the tested inputs and captured responses.
Repeated trials reveal observed variability, and scheduled checks help identify changes against reviewed expectations. Policy checks and human approval support controls at connected action boundaries. The tested coverage and actual enforcement boundaries still determine what these checks can demonstrate.
You can add SDK/API checks to an existing application, which enforces the returned decisions. Runtime requires a compatible workload deployed and operated in your environment. For supported coding agents, the passive observer collects recorded activity without blocking commands. Integration work depends on the connection you choose.
Your application stays in your environment when using the SDK or API. You deploy and operate compatible Runtime workloads in your infrastructure. The passive observer runs on your development machine. Connected hosted evaluation and evidence services receive the activity your integration submits.
Hosted evaluation receives the content your connection submits. Observer capture, redaction and retention depend on configuration; redaction does not guarantee all sensitive content is removed. We’ll review your data flow and handling requirements when scoping the connection.
Failure handling depends on the connection and its settings. For actions that require authorization, use fail-closed handling so an unavailable check cannot permit execution. Validate service failures, invalid responses and timeouts, including what happens when human approval expires.
Network time, policy complexity and enabled checks affect response time. Measure typical and slow requests with your own workload. Human review deliberately adds a wait; align application timeouts with the approval window.
Yes. Run reviewed assurance suites on a schedule or from your release workflow. Compare against a chosen baseline and send configured alerts through email, Slack or webhooks.
A passing assessment shows that the tested scenarios met their assertions. Approval authorizes an action; it does not prove execution. Recorded outcomes show what the connected application reports happened. These records support review without establishing universal correctness or certifying compliance.
Bring one workflow, the actions it can take and the rules you want to test or enforce. We’ll map the connection, data flow, approval needs and evidence. Agree assessment scope, success criteria, volume, support and pricing before activation.
THE NEXT STEP IS ONE AGENT
See how DriftGard fits your agents, your rules and your team.