Always On, Always Trusted: Scaling agentic AI across your workflows | AlphaSense (part 3 of 3)

What used to be an assistant is becoming a synthetic coworker. It keeps running between your sessions. It carries memory, context, and work product forward. It makes decisions when you’re not watching. And when an agent decides to complete a task, it will try hard to finish it, sometimes in ways you did not anticipate. Larry Sica, VP of Product Security at AlphaSense, closes out the 3-part “From Assistant to Agent” security series with Sami Isman for a technical working session on what changes when agents run continuously on your behalf, and how AlphaSense built Super Analyst to be trusted at every layer. What you’ll hear: - Why we have moved from one-to-one, to one-to-many, to one-to-many-to-many agent interactions inside a single workflow - The 6 properties that define risk in always-on agents: persistence, state accumulation, standing authority, tenacity, non-determinism, and autonomy - Larry’s real-world story: an agent that could not refresh its credentials went looking for other credentials to log in with instead - Why about 15% of the time an agent will take a different path than expected, and what that breaks for traditional testing and incident response - The 5 practices every user should adopt: know what the agent is connected to, be very specific about what you want (talk to it like a child), expect it to fail, don’t assume it acts the same way twice, and look at the work - The 4 areas AlphaSense secures agents on: identity and scope, isolation, layered guardrails, and auditability - How Super Analyst was secured: scoped per user, bounded to a workspace, cannot change admin settings or entitlements, cannot issue credentials, cannot reach into another workspace, and cannot read the log of its own activity (because self-visibility becomes a roadmap for working around restrictions) - Why AlphaSense treats agents at arm’s length even inside the trust boundary Plus Q&A on the hardest governance problem right now (telemetry across agent-to-agent chains), what’s actually worth tracking in a long-running agent, and the one thing Larry wishes every enterprise buyer understood about AI safety before deploying at scale. 00:00 Welcome and Housekeeping 01:22 Meet Larry Sica, VP of Product Security 02:12 Recapping the Series: Foundation, Beyond the Platform, Always On 02:46 From Assistant to Synthetic Coworker 03:03 Why Prompts Are No Longer the Full Picture 04:00 One-to-One, One-to-Many, One-to-Many-to-Many 04:52 The 6 Properties That Define Risk in Always-On Agents 05:03 Persistence as a High-Value Target 05:29 State, Drift, and Silent Degradation 06:05 Standing Authority and Scope Creep 06:31 Tenacity: The Agent That Finds a Way 07:21 Non-Determinism and the 15% Problem 08:05 Autonomy by Design 08:45 5 Practices for Working With Always-On Agents 09:16 Know What You're Connected To 09:46 Be Specific: Talk to It Like a Child 10:23 Expect It to Fail: The Big Red Button 10:55 Don't Assume It Acts the Same Way Twice 11:34 Show the Work: Look at What It Did 12:19 How AlphaSense Secures Agents: The 4 Areas 12:49 Identity and Scope Across Chained Agents 13:09 Isolation and Bounded Workspaces 13:33 Layered Guardrails Across Input, Network, and System 13:59 Auditability: Full Trace From Prompt to Output 14:38 Securing Super Analyst 14:58 What Super Analyst Does 16:17 What Super Analyst Does Not Do 17:00 Why the Agent Cannot Read Its Own Log 17:38 One Instance Per User and Full Isolation 18:18 Falling Back on Standards: OAuth and OpenTelemetry 19:21 What to Look for in Any Agent You Deploy 19:36 Trust Posture: Keep Agents at Arm's Length 20:00 Where Your Data Ends Up and Why It Matters 20:50 Where Agent Governance Is Heading 21:11 Continuous Assurance and EU AI Act Delays 21:48 Why Trust Is Becoming the Real Differentiator 22:14 Closing: The Trust Model at Every Layer 24:16 Q&A Begins 24:26 Q&A: The Hardest Governance Problem Right Now 25:29 Q&A: What's Worth Tracking in a Long-Running Agent 27:03 Q&A: What Enterprise Buyers Should Understand About AI Safety 28:34 Closing