Azirella
← Back to Autonomy

Cross-boundary decisions

Agent Negotiation

In real enterprise operations, no single person or system controls everything. A demand planner can't unilaterally increase production, that's the plant manager's domain. A procurement analyst can't redirect logistics, that's the logistics team's call. Meaningful responses across the six decision domains cross multiple authority boundaries.

Traditional systems handle this with emails, meetings, and phone calls. A demand spike triggers a week of cross-functional coordination before anyone acts. By then, the opportunity, or the crisis, has moved on.

Autonomy solves this with Agent-to-Agent authorization, autonomous agents negotiate cross-boundary decisions in seconds, under the AIIO operating model with full transparency and human override at every step.

How Agent Negotiation Works

From request to cross-functional resolution in seconds

NATURAL LANGUAGE REQUEST "500 C900 bikes to Detroit in 2 weeks" STRATEGY GENERATION 3 candidate strategies evaluated in parallel AUTHORITY BOUNDARY CHECK Partition actions: within authority vs. cross-boundary WITHIN AUTHORITY Execute immediately e.g., Raise order priority to P1 CROSS-BOUNDARY Requires agent-to-agent authorization e.g., Production increase, PO expedite AGENT-TO-AGENT NEGOTIATION Plant Agent AUTHORIZE ✓ Procurement Agent COUNTER-OFFER: 5-day EXECUTE & RECORD Actions applied · Full audit trail in Decision Stream Human can Inspect or Override at any time < 1s 3-5s < 1s 2-4s < 1s

A Real Example: The Rush Order

From natural language request to cross-functional execution in 15 seconds

"Bigmart just called, they need 500 C900 bikes delivered to Detroit in 2 weeks. This is a new fleet deal we can't lose."

, Demand planner, via Azirella Assistant
1
Compound Intent Detection < 1 second

The system recognizes this as both a demand signal (a new customer order) and an implicit directive (do whatever it takes to fulfill it). It creates the order and checks feasibility.

Current plan can only promise 320 of 500 units. Shortfall: 180 units.
2
Strategy Generation 3-5 seconds

The AI strategist generates three candidate resolution strategies:

A
Reprioritize ATP, Steal allocation from lower-priority orders → 92% fill
B
Increase Production, Rush manufacturing order at Plant 1 for 80 additional units → 95% fill
C
Reprioritize + Expedite, Raise priority AND expedite component POs → 100% fill ✓
3
Authority Boundary Check < 1 second

Strategy C has three actions. The demand planner's authority is checked for each:

Raise order priority, Within authority. Executed immediately.
Add 80-unit production order, Cross-boundary. Needs Plant agent authorization.
Expedite component PO, Cross-boundary. Needs Procurement agent authorization.
4
Agent-to-Agent Negotiation 2-4 seconds

Cross-boundary actions are sent to the owning agents for evaluation. Each agent checks its current domain state and responds:

Plant Agent → AUTHORIZE
"Plant capacity at 78%, 22% spare capacity sufficient for 80 units."
Procurement Agent → COUNTER-OFFER
"4 active expedites out of 5 allowed. Can offer 5-day delivery instead of 3-day."
5
Execution < 1 second
Priority raised: Order updated to P1, ATP re-consumption makes 320 units available immediately
Production order created: 80 units of C900 at Plant 1, authorized by Plant agent
PO expedited: Component delivery moved to 5-day (tweaked from 3-day by Procurement), counter-offer accepted
6
Decision Stream recorded

The entire decision, all three strategies evaluated, the winner selected, the agent-to-agent conversation, the counter-offer from Procurement, is recorded as a single decision in the Decision Stream. Any stakeholder can Inspect the reasoning or Override any action.

~15 seconds
What takes 3-5 days of cross-functional meetings in traditional planning

The AIIO Decision Model

Every decision in Autonomy follows the four AIIO states. This is the governance model that makes autonomous operation safe, agents decide first (speed), but humans always have the final say (governance).

Automate

System auto-selects the best strategy and executes within-authority actions. No human delay for routine decisions.

Inform

Decision surfaced to the Decision Stream with full reasoning, strategy comparison, and agent-to-agent conversation.

Inspect

Planner reviews the comparison table, authority boundaries crossed, and agent reasoning. Full audit trail.

Override

Human selects a different strategy or rejects an action. The override reason feeds back into agent training, teaching agents which overrides improve outcomes.

This mirrors Kahneman's insight about fast and slow thinking applied to organizations. The agents are System 1, fast, pattern-matched, handling routine decisions automatically. Human planners are System 2, slow, deliberate, activated when something looks wrong. The AIIO model ensures System 2 always has visibility and veto power, without requiring it to process every decision.

Authority Boundaries

Every agent has three categories of actions. These boundaries mirror how real organizations work, a demand planner can adjust priorities (their domain), but requesting a production increase crosses into the plant manager's domain and requires authorization.

Unilateral
Can execute without asking anyone. Example: ATP analyst raises order priority.
Requires Authorization
Must get approval from the domain owner. Example: Demand planner requests production increase.
Forbidden
Cannot request under any circumstances. Example: Analyst overrides strategic policy parameters.

15 agent roles span the enterprise: Sales/ATP, Supply, Allocation, Logistics, Inventory, Plant, Quality, Maintenance, Procurement, Supplier, Channel, Demand, Finance, Service, and Risk. The authority boundary map is exhaustive, every action type is mapped to exactly one domain owner. Unknown actions default to requires-authorization (pessimistic safety).

How Agents Respond

When an agent receives an authorization request, it evaluates against its current domain state, inventory levels, capacity utilization, active commitments, and responds:

AUTHORIZE

Feasible, no contention. Action executed as-is.

COUNTER-OFFER

Feasible with modifications. "I can do 60 units instead of 80 without overtime."

DENY

Infeasible or constraint violation. Action skipped, reason recorded.

ESCALATE

Agent uncertain. Pushed to Decision Stream for human Inspection.

Counter-offers are the most common outcome, agents rarely say "no" outright. Instead they negotiate: "I can't do 80 units at standard cost, but I can do 60 without overtime." This mirrors how real cross-functional teams work, but at machine speed.

The Agentic Operating Model

This is what the agentic inversion looks like in practice. Agents own decisions by default. Humans Override with reasoning captured. The more decisions flow through the system, the better the agents become, and the less human effort is required for routine coordination.

15s
vs. 3-5 days of meetings
100%
audit trail for every decision
15
agent roles with explicit authority boundaries

The judgment layer, knowing when to Override and why, becomes the organization's competitive advantage. Override reasons feed back into the learning flywheel, teaching agents which human judgments consistently improve outcomes. Over time, the agents internalize the organization's decision culture.

See Agent Negotiation Live

Watch agents resolve a cross-functional crisis in real-time.