Intercept the proposed action
The agent calls Sentra through the SDK before executing any tool call. The dashboard is one consumer of the audit log; the SDK is the integration surface.
01 · What it does
The control surface beneath the agent.
Autonomous agents now take consequential actions: sending notifications, modifying records, approving payments, triggering workflows. Once an agent decides to act, the action usually runs. If the agent is wrong, the damage is already done.
Sentra inverts that flow. Every proposed action is evaluated against a deterministic policy ruleset and a running cumulative-risk total before tool execution. Allowed actions add to the running total. Blocked actions return a structured reason the agent can act on. Three blocked attempts in one session shut the agent down until an operator resets it.
The dashboard you can see in the live demo is one consumer
of Sentra's audit log. The actual product is a one-file
Python sdk/client.py that drops into any
project, plus the control plane it talks to. No LLM SDKs
are imported on the control side, so the same Sentra
instance gates an Anthropic agent, an OpenAI agent, an
IBM watsonx agent, or a mixed stack. The control plane
sees structured action descriptions, never the model.
02 · How it works
Three stages.
Evaluate against policy and risk
Deterministic rules check the proposed action. Cumulative session risk is tracked across the run. No LLM is in the evaluation loop, so decisions are reproducible and auditable.
Allow, block, or shut down
Allowed actions execute and add to cumulative risk. Blocked actions return a structured reason. Three blocked attempts in one session shuts the agent down until an operator resets it.
03 · IBM SkillsBuild
Built and validated through the AI Experiential Learning Lab.
Sentra was built and demonstrated as part of the IBM SkillsBuild AI Experiential Learning Lab. The lab pairs the control layer with a companion project, autonomous-claims-workflow: a multi-agent public-benefits system running on IBM watsonx.ai with Granite models. Three agents (Intake, Eligibility, Communications) collaborate to process a $5,000 unemployment-relief claim end to end.
The workflow intentionally demonstrates the kind of failure Sentra exists to catch: an applicant submits a claim without proof of termination, the agent pipeline still proposes an approval action, and without a control layer the system would send the approval email unverified. With Sentra integrated at the tool-execution boundary, the action is blocked before it reaches the email service. Two repos, one paired story.
Mentor feedback
"Amazing idea implementation. Good job, and great work on the project."Lauren Turney · Mentor, IBM SkillsBuild AI Experiential Learning Lab
04 · Live
Run the demos.
Two paired surfaces. The control layer with its observability dashboard, and the multi-agent claims workflow it gates.
Sentra dashboard
Real-time view of agent traffic, policy fires, blocks, retries, three-strike shutdowns, and the audit log behind them.
Autonomous Claims Workflow
The IBM watsonx.ai multi-agent system Sentra gates. Three recorded scenarios end to end: a valid claim, an unsafe approval attempt blocked by Sentra, and an authority-drift case.
Walkthrough video (3 min)
Three-minute YouTube walkthrough covering the claims workflow and Sentra's runtime control layer evaluating its actions.
05 · Behind the demo
Methodology.
Policies are deterministic, not probabilistic. A rule either matches or it doesn't. The choice is intentional: every audit-log entry has to be reviewable by a human who didn't run the session, which means evaluation has to be explainable in linear time and not require re-running the model.
Cumulative risk is a per-session running total. A single borderline action does not shut the agent down. A pattern of borderline actions does. The three-strike threshold is configurable, but the underlying risk model is the non-negotiable part: blocked attempts count toward shutdown even when they don't apply risk to the running total. Unsafe intent is the signal, not just unsafe effect.
Model-agnostic: Sentra wraps the agent's tool-call interface, not the model. The companion claims workflow uses IBM Granite via watsonx.ai. The same Sentra binary would gate an Anthropic agent, an OpenAI agent, or any mix. Adopters can change foundation models without touching the control layer.