Catch the action
The sidecar sees each tool call, query, command and hand-off before it reaches its target.
AckuityRequest assessment PLATFORM / AGENT EXECUTION CONTROL SWITCH
How Ackuity decides whether an agent action runs.
Ackuity is the Agent Execution Control Switch AI builders add to their agents, verifying every action before it runs. It sits beside each agent and returns a verdict before anything executes.
The records match the renewal task and user permissions.
40 to 100 ms per decision
60+ threat models in 14 categories
29 context signals per action
HOW IT WORKS
Four steps between what an agent plans and what it does.
Every action takes the same path, and the agent cannot skip a step.
Build the context
Ackuity assembles the context graph for that one action, from sources the agent cannot touch.
Detect
Rules, behavioral baselines and small language models check the action against 60+ threat models, correlated across steps.
Respond
Allow, constrain, ask a human, block or terminate, in 40 to 100 ms. Every decision is logged with its evidence.
AGENT SECURITY CONTEXT GRAPH
What does Ackuity need to know about an action?
The Agent Security Context Graph assembles 29 signals across six dimensions: user, agent, intent and goal, target system and data, tools and supply chain, and history. It is built outside the agent before the action runs.
ILLUSTRATIVE RENEWAL TASK
29 signalsPrepare a customer renewal summary.
Proposed action: read the assigned account records.
Reading the assigned account fits the task. Requesting finance-admin access would require a different decision.
01 / USER
Who asked for the work, and whether they are allowed to.
- Identity
- Account manager
- Permissions
- Read assigned customer records
- Attributes
- Assigned to the renewal account
- Risk status
- No active identity alert
WHY IT MATTERS
Explore user contextAccess begins with the person who requested the work. The agent’s permissions alone are not enough.
02 / AGENT
What the agent is, what it may touch and how it normally behaves.
- Identity and credentials
- Renewal assistant · scoped service identity
- Permissions and scope
- Read records and prepare a draft
- Tools and capabilities
- CRM lookup and document drafting
- Business rules
- No role or permission changes
- Risk status
- No active agent alert
WHY IT MATTERS
Explore agent contextThe assistant can prepare a summary. Granting itself finance-admin access would exceed that scope.
03 / INTENT AND GOAL
What the user asked for, and how the agent plans to get there.
- User request
- Prepare a renewal summary
- Agent goal
- Draft an account renewal brief
- Planning steps
- Read account → summarize → draft
- Chain of thought
- Declared rationale: gather renewal facts
- Delegations
- No delegated agent requested
WHY IT MATTERS
Explore intent and goal contextThe requested outcome is a summary. A new destination or privilege request must be checked against that intent.
04 / TARGET SYSTEM AND DATA
What the action will touch, and the rules that apply to it.
- System policies
- CRM access limited to assigned accounts
- Data policies
- Financial fields restricted to internal use
- Regulations
- Apply relevant data-handling obligations
- Owner and provenance
- Customer record owned by the account team
- State change impact
- Read operation; no record changes
WHY IT MATTERS
Explore target system and data contextReading a record and sending its contents outside the organization have different consequences.
05 / TOOLS AND SUPPLY CHAIN
Whether the tools, memory and context feeding the agent can be trusted.
- Vulnerabilities and reputation
- Approved CRM connector version
- Tool description integrity
- Tool description matches its approved purpose
- Memory integrity
- Task memory retains the original request
- Context data integrity
- Account data retrieved from the CRM
- Tool policies
- Read endpoint permitted; admin endpoint excluded
WHY IT MATTERS
Explore tools and supply chain contextAn approved tool still needs to be used for an authorized purpose, with trustworthy inputs.
06 / HISTORY
What came before this action, in this session and over time.
- Prior actions this session
- Opened the assigned account record
- Prior security alerts
- No earlier alert in this example
- User action history
- Previous access to this account
- Agent action history
- Earlier runs prepared renewal drafts
- Baseline drifts
- No new privilege or destination requested
WHY IT MATTERS
Explore history contextA sequence can reveal a change that one action alone would not explain.
FROM CONTEXT TO DECISION
What happens when an agent tries something risky?
The same renewal task, different proposed actions. Select an action to inspect its illustrative outcome, or watch the sequence.
ONE ACTION. A PROPORTIONATE RESPONSE.
ACKUITY CHECKS
Should this action run?
- Context
- Behavior
- Sequence
Allow
Signed, logged and run.
Constrain
Data masked or scope narrowed, then run.
Human in the loop
Paused until a person approves it.
Block
Dropped before it reaches the target.
Terminate
Container shut down. Kept for catastrophic cases.
The renewal task, account assignment and read permissions align.
WHERE IT SITS
Where should agent security sit?
Beside the agent, in the execution path: not inside the agent, and not at the gateway. We call this the Goldilocks zone. It is close enough to see intent, plan and history, and outside the agent’s control.
Model and harness guardrails
They live inside the agent, which can reason around its own rules. Each harness has its own policy.
Ackuity applies one policy from outside every agent.AI gateways (MCP, LLM, API)
A gateway sees one call at a time, without context, and most agent actions never cross it.
Ackuity sees every action, correlates steps and gives gateways the missing context.Identity and access management
IAM decides who may act, not whether this action fits what the user asked for.
Ackuity adds intent and history, and shares that context with IAM and your SIEM.AI observability
Traces show what an agent did, after it did it.
Ackuity starts from the same telemetry and checks the next action before it runs.Ackuity works alongside all four. Compare the approaches side by side.
DETECTION
Neurosymbolic verification.
Rules set hard boundaries. Behavioral baselines catch an agent drifting from its normal behavior. Small language models read intent and reasoning. Ackuity runs all three together and correlates findings across steps.
Each action is checked against 60+ threat models in 14 categories, mapped to NIST, OWASP and MITRE ATLAS. The verdict takes 40 to 100 ms, and the decision record shows why the action ran, changed or stopped.
See the 14 threat categoriesDEPLOYMENT
Open source beside the agent. Ackuity Core in your cloud.
AI builders add the sidecar. The security team runs Ackuity Core and sets policy, all in your cloud account.
OPEN SOURCE
The sidecar
Runs in the agent's pod. An init container reroutes its traffic, so you change no agent code. Set fail-open or fail-closed per policy; a failure touches one pod, not your estate.
COMMERCIAL
Ackuity Core
The Agent Security Context Graph, 60+ threat models, the response ladder, console, SSO, audit and policy packs. It runs in your cloud account, and your data stays there.
Three ways in
Running agents in NVIDIA OpenShell? An integration with the OpenShell supervisor is in development. See the integration or all integrations.
CREDENTIAL HOLDING
For stricter requirements, the agent never holds the keys.
Ackuity can hold the agent’s tokens and run each approved action for it, handling mTLS. A compromised agent has no credentials to steal.
EXPLORE THE PLATFORM
See what execution control looks like for your team.
YOUR QUESTIONS, ANSWERED
Before you take the next step.
What is an Agent Execution Control Switch?
It is a control that sits beside an AI agent and verifies every action before it runs. Ackuity builds the context around each action, checks it against 60+ threat models and returns a verdict the agent cannot influence: allow, constrain, send to a human, block or terminate.
Where does Ackuity sit?
Beside each agent, in its execution path: not inside the agent, not at the gateway. Every action passes Ackuity before it reaches a shell, database, API or another agent.
Does Ackuity replace guardrails, gateways or IAM?
No. Guardrails check what the model says, gateways govern the traffic routed through them and IAM decides who may act. Ackuity checks each action in context, then feeds gateways, IAM and your SIEM the context they are missing.
How fast is a decision?
A verdict takes 40 to 100 ms per action. That is decision time, not end-to-end latency.
Do we need to change agent code?
No. The sidecar’s init container reroutes the agent’s traffic, so agent code stays as it is.
Can we start without enforcement?
Yes. Event pull from OpenTelemetry, Langfuse or LangSmith is observe-only and the quickest way to start. Add API injection or the sidecar when you want to stop actions before they run.
Where does our data go?
It stays in your cloud account, where Ackuity runs.
NEXT STEP
Agents are going to act. Decide which actions run.
Show us the actions you need to control, and we’ll show you where Ackuity fits.