The Cognitive Revolution

Vijoy Pandey on why two AI agents still cannot book you dinner

Vijoy Pandey· SVP & GM of Outshift at Cisco
·~90 min·English·The Cognitive Revolution
AgentsAI InfrastructureOpen SourceAI Safety
TL;DR

Six months after multi-agent systems went into production, Cisco's Vijoy Pandey open-sources the protocols that let agents align and negotiate, and argues the hardest unsolved problem is still getting two personal agents to simply agree on dinner.

01The Release

Accord Before, Concord During

Outshift has open-sourced the layer that was only a vision last time: <strong>Accord aligns agents on intent and terms before the work, Concord coordinates and negotiates during it</strong>, with Scope watching underneath and an eval protocol still to come.

engineering assumes 100,000 users in total whereas the customer really meant 100,000 concurrent users. There is a grounding problem here. There is a discovery problem here. There is a term alignment problem here.

— Vijoy Pandey, The Cognitive Revolution
Key Insight
The release is deliberately parasitic on what already exists: the protocols ride above MCP and A2A and drop in as skills, so adopting them costs an agent team almost nothing. That lowers the real barrier to alignment tooling, which was never the idea but the friction of rewiring a working stack.
02Architecture

Don't Decouple the Identity

Everything in an agent is coming apart into swappable parts, but Pandey draws one hard line: <strong>the model and the harness are interchangeable; the identity, persona and memory are not.</strong>

The harness is decoupled. Model is decoupled. But the identity is immutable and is mine or my agents.

— Vijoy Pandey, The Cognitive Revolution
Key Insight
Making identity the fixed point reframes lock-in as a data problem, not a model problem: whoever owns the persona and memory owns the relationship, no matter whose model is running underneath. It is the same move that let phone numbers survive carriers, applied to agents.
03The Proof It's Unsolved

Two Agents, One Dinner, No Reservation

The cleanest evidence that collaboration is the unsolved layer is tiny: <strong>Pandey's agent and his wife's agent still cannot book a dinner reservation on their own</strong>, blocked by the model, the permission-hungry harness, and the absence of any way to compromise.

our agents cannot do this autonomously, which is ridiculous.

— Vijoy Pandey, The Cognitive Revolution
Key Insight
The restaurant case is more damning than any enterprise demo because the scope is trivial and the failure is total. If two agents who each fully represent a willing human cannot negotiate a dinner, the gap is not scale or compute; it is that the give-and-take layer is still missing in practice, which is exactly what Pandey is now shipping Accord and Concord to supply.
04The Taxonomy

Guardian Angels and Accelerators

Borrowing a line from Raj Reddy, Pandey sorts helper agents into two jobs: <strong>guardian angels that keep you from doing what you should not, and cognitive accelerators that speed up what you already know</strong> while the human keeps the judgment call.

Gats make sure that you are guard railed and you're not doing something that you should not be doing.

— Vijoy Pandey, The Cognitive Revolution
Key Insight
Splitting guardrails from acceleration is also a safety argument: the component that restrains an agent is a separate agent with its own job, not a prompt bolted onto the one trying to get work done. Keeping the brake off the accelerator pedal is why he thinks the stack can move fast and stay bounded.
05The Bet

Latent Space, Rolled Out Like a Drug

Pandey thinks agents talking in high-dimensional latent space instead of words is inevitable and worth chasing for hard problems like drug discovery, but insists on the discipline of a clinical trial: <strong>test, canary, then deploy, and skip no step.</strong>

we should deploy it in a canary way so that we understand it in the same way we introduce a new pharmaceutical drug in the environment.

— Vijoy Pandey, The Cognitive Revolution
Key Insight
The drug framing quietly concedes the interpretability worry rather than dismissing it: you do not canary-test something you can fully read. By accepting that latent-space messages are partly unreadable, he moves the safety question from can we understand it to can we contain and roll back a failure, which is the question operations teams already know how to answer.
06The Long View

We're Already the Cow

Asked about Harari's warning that humans will end up like cattle ruled by systems they can't grasp, Pandey's answer is that <strong>we already are, and what keeps it workable is specialists who understand each piece plus a record that keeps the whole accountable.</strong>

software is running trading today and I might be the cow even today.

— Vijoy Pandey, The Cognitive Revolution
Key Insight
His comfort rests on an assumption worth naming: that for every opaque system there remains some human who genuinely understands it. The cow analogy bites hardest precisely where that stops being true, which is why he leans so hard on Scope as a system of record. The ledger is his hedge against a world where no single person understands the whole.