Cisco's AI Agent Alignment: Scope, Accord, Concord

A dashboard can glow green while nothing gets done. That is the finding Cisco's Outshift team kept running into as it built out multi-agent systems this year: agents checking in, tasks moving, status lights all clear, and the actual job finished exactly zero percent of the time. In one internal test, a simple travel-booking system reported 66% uptime (38:15), meaning the agents were running most of the time, while completing none of what it was asked to do. The system was not broken in any way a server log would catch. It was broken in a way only a conversation could catch: the agents had quietly stopped meaning the same thing.

That is the subject of this conversation with Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, the company's incubation unit for emerging technology. Pandey has spent the past year watching AI agents, programs built on large language models that can use tools and take actions on their own, move through three stages. First they were isolated, solo agents doing one job. Then they became islands, several agents working together inside one company's walls. Now, in his words, they are connected: agents from different vendors, different cloud providers, and different companies working on a single outcome together (07:02). Pandey calls the pace of change 'fly years,' not dog years, because, as he puts it, 'I like to call it fly years, it's not even dog years. It's definitely not human years. It's, it's definitely fly years because it's happening, things are happening every hour' (06:37).

Connection turned out to be a different thing from understanding. Outshift looked at public traces from ChatDev, an open multi-agent coding environment, and found that 100% of the sessions it examined contained what Pandey calls semantic failures and cognitive misalignment (36:31): agents working side by side without agreeing on what they were actually working toward. In Outshift's own five-agent deployments, the arithmetic of this problem turned out to be brutal. One uncorrected misunderstanding between agents dropped a system's accuracy from 99% to 98%, tolerable. A second dropped it to 81%, still survivable. A third dropped it to 16%, close to a coin flip (37:05).

"We are calling these kinds of failures silent failures with really loud consequences because you don't see them today, but they have really loud consequences." Vijoy Pandey [39:16]

One of the clearest working examples sits at Swisscom, the Swiss telecom operator. The company wanted a system that could reconfigure its data centers during hardware and software migrations with zero errors, a problem too large for any one engineer to hold in their head (10:56). Outshift's answer paired a formal, rule-based model of the network, the parts of the system that cannot be violated, with five agents that explore possible configurations within those rules (13:59), aiming for a 100% error-free result (12:57). Pandey compares it to playing chess against an opponent who has already mapped every legal response: there is no winning, but there is also no illegal move.

Why Agents Need Each Other, and Why That Breaks Down

Pandey's explanation for why these systems need multiple agents, rather than one very capable one, leans on an analogy from human history. For hundreds of thousands of years, individual humans got smarter without getting more effective as a group, because insight died with the person who had it. About 70,000 years ago, language arrived and changed that (28:12). It let people share intent and set a common goal, pass knowledge across generations instead of relearning it each time, and eventually reason together instead of alone. Pandey thinks AI agents are walking a compressed version of that same path: individually capable, but not yet able to reason collectively. 'We've realized that connectivity is not the same as cognition. What we need to move towards is shared cognition and shared reasoning' (29:50).

The example he returns to is a mundane one. An engineering team is told to build a feature for '100,000 users.' Sales means it as a target market size. Engineering reads it as total users ever. The customer meant 100,000 concurrent users at once (46:17). Nobody lied. The term was simply never pinned down, and the ambiguity compounds through every handoff that follows, the same way it would in a human team working across sales, legal, finance, and engineering on a multimillion-dollar deal (18:40).

To catch and fix this, Outshift has open-sourced three tools, released as skills that plug into existing agent environments like Claude Code and Codex rather than requiring anyone to rebuild their stack. Scope is an observability layer, built on OpenTelemetry, an open standard for recording what software is doing, that reads a system's own traces and flags where meaning has drifted, the way lane-departure software in a car warns a driver before a crash rather than after (38:52). Accord gets agents to agree on shared intent and ambiguous terms before work starts (42:39). Concord lets them renegotiate as goals shift while work is underway (48:01). A fourth piece, meant to verify that a finished task actually matches what was agreed, is still being built (48:25).

One Identity, Many Harnesses

The conversation's most concrete personal detail is Pandey's own AI agent, named Arnold Layne after a 1967 Pink Floyd song. The naming makes an architectural point as much as a playful one: Pandey insists that an agent's identity, personality, and memory should stay fixed while the harness, the software wrapper that manages tool access and guardrails, and the underlying model churn underneath it. Arnold has run on several different model providers over time, moving as costs and capabilities shifted, while remaining, in Pandey's telling, the same entity (65:55). The reasoning is practical: tying identity to one vendor's model or interface becomes a trap once that vendor's advantage fades.

Even this setup does not solve the coordination problem. Pandey's wife has her own agent, named Lyra Belacqua after a character in Philip Pullman's novel His Dark Materials. The two agents still cannot book a dinner reservation together without a human stepping in, because they disagree on restaurant preferences and because the harness keeps asking for permission at each step, which defeats the point. 'If I'm going to give you permission every single time, I might as well just do it myself,' Pandey says (69:40). If two agents representing two people in the same household cannot negotiate a restaurant booking, the gap between today's agents and a fully autonomous enterprise deal desk looks substantial.

Pandey frames the near future around two agent types, a framing he attributes to Carnegie Mellon computer scientist Raj Reddy: AI should be either a guardian angel, keeping a system inside its guardrails, or a cognitive accelerator, speeding up work that is already understood (61:24). Judgment calls, the ones that require taste rather than data, still belong to a human sitting 'on the loop' rather than being fully replaced (58:15), though Pandey expects some of that judgment to eventually be learned by what he calls a cognitive accelerator.

Toward the close, Pandey takes up historian Yuval Noah Harari's recent argument that humans may end up like cows: governed by systems, in Harari's case financial markets, that benefit us but that we cannot fully understand (81:09). Pandey accepts the comparison, then relocates it. He already does not understand high-frequency trading software, yet it runs markets he participates in every day (82:20). What matters, in his view, is not that every human understands every system, but that someone does, and that the system keeps a full record of what it did and why, so it can be audited and corrected when it drifts (84:46). That same logic, one layer down, is the case for Scope, Accord, and Concord. Somebody, or something, should always be watching the seam where one agent's meaning meets another's, with a full record kept for when it drifts.