Ethics in agentic AI

The three analogies LLMs = consultants (give advice, don’t implement). LLMs with functions = assistants with tools (can use tools when told). Agents = Trusted employees (you give them goals, they figure it out)………..yes *

As AI systems evolve from passive consultants to autonomous goal-driven agents acting as trusted digital employees, where should ethical reasoning, control, and responsibility reside within agentic AI systems?*

Is ethical governance primarily a design-time responsibility of developers, a usage-time responsibility of users, or a dynamic shared responsibility enforced through agent-level constraints and oversight mechanisms?

Ethical governance in agentic AI should be shared. Developers must design strong safeguards and constraints, users must set responsible goals and provide oversight, and the system should include continuous monitoring and accountability mechanisms. It cannot rely on only one side.

I’m not placing the onus on any single side. I’m seeking perspectives on the possible consequences of unethical practices and, more importantly, on reasonable ways to mitigate them while designing tools, services, or products.

We already have strong guiding frameworks—such as Google’s own AI Principles ( AI Principles — Google AI )—and my intent is to understand how these principles can be explored more deeply and made more actionable for a broader community. You can eloborate more over this or extend this thread …as per your choice

FROM MY POV failure of complying with SDLC by agent are creating major issue because It must, but it often doesn’t.

the SDLC (Software Development Life Cycle) is a predictable map. With Agentic AI, the “software” is no longer a static set of instructions—it is a dynamic entity that plans and executes its own steps. This creates a massive friction point between standard engineering and the unpredictable nature of agents.

WHY AGENTIC AI DISOBEY SDLC

In a normal SDLC, we write unit tests for specific functions. You cannot “unit test” an agent’s reasoning process easily because it might solve a problem three different ways in three different runs.

• Once we deploy a standard app, it stays the same. An agentic system “evolves” as it interacts with tools and memory. This makes “version control” extremely difficult—are you versioning the code, the model, or the agent’s learned state?

• Requirements usually define what a system does. For agents, we define a goal, and the agent decides the how. This flips the “Design” phase of the SDLC on its head

**

ONE OF THE MAIN PRINCIPAL OF RESPONSIBLE AI TRANSPARENCY GOES FOR A Toss AND CREATES COMPLIANCE GAP as we never know will agentic ai will jump the gun break another traditional rule.

**

Solution
Define clear boundaries for agents , give them no scope to jump a gun , give specific clear set of do’s and dont’s .

REDEFINE SDLC : as SDLC do not now ends at deployment rather it must be a continuous process , a must always require evaluation as agent always evolve and often find changes in goal definition

Red Teaming : it should be common and continuous practice force agents to break set rules before it sees end customer , We can use agentic ai to trick agents in to breaking rule

HIL ( Human in loop ) : CREATE CHECKPOINTS IN CODES for ex for high risk actions the agent’s SDLC must force it to pause and wait for a human "API call

agents must be deployed in “Shadow Mode” first—they must watch real data and “suggest” actions without actually executing them, allowing developers to verify the SDLC logic in the real world.

SDLC IS A main artery of RESPONSIBLE AI so SDLC MUST EVOLVE AND MUST BE FORCED STRICTLY .

Its time to realise Agentic AI cannot survive without an SDLC, or it becomes “Shadow AI”—untraceable, unfixable, and a massive compliance liability if we are Serious about RESPONSIBLE AI FOR ALL

good insights and its worthwhile to mention that “WHY AGENTIC AI DISOBEY SDLC” …and moreover the versioning cant be guranteed about this and how much more time it will take in future too and this discussion goes on …. thanks keep on continue this threads for more to add by others