Opinion | When Software Becomes A Counterparty

Aditya Vikram Kashyap

As AI shifts from questions to actions, software is becoming a delegated actor. The next challenge is not intelligence, but proving who authorised it and who is accountable

The transition from generative to agentic AI is usually described as software that answers becoming software that acts. That framing is accurate and insufficient. The deeper transition is from software as tool to software as delegated actor, and it is institutional before it is technical. When a system can hold state, select tools, invoke other services, and trigger transactions under delegated authority, it begins to occupy operational positions that were previously held only by people and organisations. The other side of those interactions inherits questions we normally ask about counterparties. Who is this? Whose authority does it carry? What can it commit? Who answers for what it does?

I use “counterparty" as an architectural metaphor, not a legal claim, and I am not suggesting that AI agents possess legal personality. The idea is narrower: software is starting to occupy positions in which other systems must evaluate its identity, authority, provenance, and accountability before accepting its actions. Our security architecture was not built for that evaluation, because it was built around a different shape of relationship entirely.

Best Deals on Kitchen Chimneys and Cooktops Under Rs. 12,000: Here Are Our Top Picks to Grab

Virat Kohli’s Intensity 'Pushes' Teammates To Raise Their Game: 'What Impressed Me More...'

Crores Spent, Machines Lie Idle: Delhi Govt Hospitals Report Hundreds Of Unused Medical Equipment To HC

Singer Sukriti Kakar Gets Engaged To Shoumik Shetty, Shares Photos From Dreamy Proposal

Authority At A Distance

The old shape was short and legible: a human authenticates, receives permissions, issues an explicit instruction, and a machine executes it. Identity, authority, judgment, and action all sat with one person, and responsibility stayed conceptually attached to that person. Automation preserved it: a batch job executes a predefined process; the process is the authorization.

Agentic systems pull those four things apart. A principal, the person or institution whose authority is ultimately being exercised, delegates an objective. An actor, the software entity currently executing, decomposes the objective, exercises judgment about pathways, and may delegate sub-tasks to further actors before anything touches an external system. The sequence of actions was not fully specified in advance. That is what makes agents useful, and it is also what creates the defining trust problem, which is not autonomy as such. It is authority being exercised at increasing distance from the principal who granted it. Call it authority distance: each hop between principal and final action risks losing fidelity between original intent, delegated scope, permitted action, and accountable party. A receiving system no longer just needs to know which actor sent a request. It needs to know which principal........

© News18