Two Agents, Two Definitions of 'Done': The Contract to Write Before You Wire Them Together
The failure mode nobody writes about: two AI agents, both running, both "up" on every dashboard — and the job quietly not getting done. On the November 11 episode of The Cognitive Revolution, Vijoy Pandey — SVP and GM o
The failure mode nobody writes about: two AI agents, both running, both "up" on every dashboard — and the job quietly not getting done.
On the November 11 episode of The Cognitive Revolution, Vijoy Pandey — SVP and GM of Outshift by Cisco, the incubation arm that has deployed multi-agent systems with Swisscom, ServiceNow, SoftServe, and Webex — described a gap that standard monitoring never shows. In Cisco's own five-agent deployments, accuracy fell from 99% with no misalignments between the agents to 16% with three. In one travel-booking test, the system was mostly up and connected while completing zero tasks. And in an analysis of public ChatDev sessions, Cisco reported semantic or cognitive-misalignment failures in 100% of the sessions it examined.
"Connectivity is not the same as cognition," Pandey told host Nathan Labenz. "What we need to move towards is shared cognition and shared reasoning."
His diagnosis: the industry measures whether agents are up and connected, while agents can be connected and still not understand each other. The value, he argued, is not in any one agent — "the value is actually in the handoff that's happening between these agents."
The failure isn't capability. It's vocabulary.
Most multi-agent advice focuses on architecture: which agent does what, who orchestrates, how they connect. The failure Pandey described is duller and more dangerous: two competent agents that mean different things by the same word.
At small-business scale, the version looks like this:
- Your sales agent marks a job "done" when the customer accepts the quote.
- Your invoicing agent considers the job "done" when payment clears.
- Your sales agent calls someone a "customer" the moment they sign.
- Your billing system calls them a "customer" only once an invoice exists.
Both definitions are defensible. Neither agent is broken. But the moment you wire automation across that seam — sales closes the job, invoicing starts billing — you have two different meanings of the trigger word routing real money. The handoff fires early, or late, or on the wrong population, and everything looks green while it happens.
Cisco's answer was to open-source three protocols, which Outshift describes as observe, align, and validate: Scope builds knowledge graphs on agent traces to find cognitive failures; Accord has agents agree on intent, scope, key terms, and the definition of "done" before work starts; Concord has agents negotiate an outcome every one can accept when they disagree.
You do not need to install any of that. You need the idea underneath it: the definitions get written down before the wiring, not after the failure.
The handoff contract: one page, 20 minutes
Before you connect any two agents — or an agent to a system it will trigger — write five fields. One page per seam; a three-agent setup has three seams, and each gets its own page.
1. Name the pair. "Sales agent → invoicing agent." Every contract covers exactly two systems. If a diagram is forming in your head, you're overcomplicating it — the contract is pairwise, not a hub.
2. List the words that cross the seam. Not every word — the load-bearing nouns, usually five or fewer: done, customer, paid, urgent, late. Pull the actual messages that will pass between the two systems and highlight the words both sides will act on.
3. Write each side's current definition. One line each, no editing, no fixing yet. Done (sales agent): quote accepted, kickoff email sent. Done (invoicing agent): invoice issued, payment reconciled. The differences are usually obvious the moment they're on paper — which is exactly why they need to be on paper.
4. Negotiate the shared definition. Where the two sides differ, write the one definition both will use — and make it checkable. "Done = quote accepted AND kickoff email sent AND invoice issued" is a definition an agent can test. "Done = the customer is taken care of" is a vibe. This is Accord's job shrunk to a paragraph.
5. Write the conflict rule. When a message arrives that doesn't fit the shared definitions, the receiving agent stops and asks a human. Never improvise a meaning. This is Concord's negotiation step shrunk to its SMB version: escalate, don't guess. The 99%-to-16% collapse came from misalignments that went uncorrected — the contract's job is to catch the collision while it's still a question, not a loss.
Then run one real past case through the seam before you turn the wiring on — the trickiest one from last month. If the two agents don't classify it identically against the contract, the contract isn't finished.
Where this fails (the honest list)
- A contract fixes collisions, not competence. Perfect definitions make two agents agree; they don't make either one good. If the model writes a bad email, the contract just ensures both agents agree it was sent.
- Sometimes the answer is one agent. If a single agent can do the whole job end to end, do that and skip the seam entirely. Contracts are for handoffs you actually need — don't create a second agent just to have something to wire.
- The numbers are a warning, not a benchmark. The 99%-to-16% collapse and the 100% ChatDev figure come from Cisco's own deployments and its analysis of public sessions. They establish the failure mode exists, not your odds of hitting it.
- Even the builders haven't finished. Pandey reported on the same episode that his own personal agent and his wife's — each carrying one identity across several models — still struggle to coordinate on their humans' behalf, down to booking a shared dinner. If the executive open-sourcing alignment protocols is still working on it, a one-page contract doesn't put you behind. It puts you ahead of most of the market.
Where to start
You can't write a handoff contract for agents you don't know you're running. Shadow usage means most small businesses already have two AI tools touching the same customers without anyone writing down what happens between them — one paid by the company, one installed by an employee, no shared definitions anywhere.
That's the first thing the AI Clarity Sprint maps: every AI touchpoint in your business — paid tools, AI features inside software you already own, and the employee shadow usage nobody approved — in two weeks, for one fixed fee ($1,495 CAD). You finish with the complete inventory, the seams worth writing contracts for, and a ranked plan for what to do first. The cheapest place to fix a misalignment is before the wiring. The second cheapest is the week you find the seam.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.