Goals and plans
How agents translate a goal into a sequence of useful steps.
Research direction
Agentic Decision Systems
Research into how AI agents decide, act, use tools, recover from errors, and continue work across steps.
AI agents need decision systems
AI agents are not useful only because they can answer questions. They become useful when they can decide what to do next, choose the right tools, act under constraints, inspect results, recover from errors, and continue work over time.
Agentic Decision Systems is inAi's research direction for that decision layer. We study how agents move from goals to plans, from plans to actions, from actions to verification, and from failures to repair - without confusing autonomy with uncontrolled behavior.
Agents need a decision layer.
A chatbot can respond to a prompt. An agentic system has to do more. It has to decide what matters, choose a path, call tools, inspect results, remember state, recover when something fails, and know when a human or another system should review the result.
That decision layer is where agentic systems become useful - and where they become hard to design. Too much freedom can create risk. Too much control can make the system useless. The research question is how to design agents that can act, continue, and repair while staying understandable and bounded by the context.
What this direction studies
This research direction studies the parts of agentic systems that sit between model capability and real execution.
Tool use
How agents choose tools, call APIs, inspect results, and decide what to do next.
State and memory
How systems preserve context, previous decisions, workflow progress, and recoverable state.
Verification and repair
How agents check outcomes, detect failures, retry safely, and recover from bad states.
Autonomy and review
How systems decide when to act freely, when to ask for approval, and when to escalate.
Coordination
How multiple steps, tools, agents, or human reviewers fit into one decision loop.
Part of the systems view of intelligence.
inAi's AGI stance is that general intelligence may emerge from systems of models, agents, tools, memory, perception, execution, coordination, feedback, and environment - not from one isolated model alone.
Agentic Decision Systems studies one core part of that view: how agents decide and act inside a system. If models provide capability, agents organize that capability into action. They connect goals to tools, state to memory, autonomy to review, and execution to feedback.
From research to agent-operable software.
This research direction also informs Products for Agents. If agents become software operators, they will need products built for the way agents work: clear tools, readable instructions, state, memory, structured outputs, permissions, recovery paths, and review where the task requires it.
Agentic Decision Systems studies the decision patterns behind that product category. Products for Agents turns some of those ideas into software surfaces agents can use.
Research outputs
This collection will gather notes, protocols, essays, and research outputs related to agentic decision systems as they are prepared for publication.
For launch, the collection lists only the deeper overview and the ADA-1 protocol frame retained for review. Additional outputs should appear only when they are prepared for publication.