Agentic Decision Systems
Agentic systems do more than answer. They decide what to do next, whether to gather more information or act, which tool to use, what state to preserve, how to inspect the result, and whether to continue, repair, ask for review, or stop.
At inAi, Agentic Decision Systems is the research direction for that decision layer. It studies how model capability becomes continued action inside a system—and what makes that action useful, adaptable, recoverable, and understandable in a particular environment.
The direction does not begin from the assumption that more autonomy is always better. It asks which decisions can be delegated, which need constraints or review, and how an agent should behave when information, tools, or outcomes are uncertain.
Why decisions are the difficult part
A chatbot can stop after producing a response. An agentic system has to manage a sequence of decisions whose consequences persist. Calling the wrong tool can change the environment. Preserving the wrong memory can distort later work. A superficially successful outcome can hide unnecessary steps, policy violations, wasted effort, or a failure that will reappear when conditions change.
That is why the central question is not simply whether an agent completed a task. We also need to understand how it interpreted the goal, what evidence it gathered, why it selected a particular action, how it responded to feedback, whether it recognized uncertainty, and whether it could recover without losing accountability.
Agentic decision-making is therefore a systems problem. It depends on the model, but also on the tools available, the state of the environment, the quality of memory, the structure of permissions, the verification process, and the rules that determine when the system may continue on its own.
The agentic decision loop
From verification, the system must choose among four paths:
The result of each path changes the state of the system and becomes part of the next decision. An agentic decision system is therefore not a fixed pipeline. It is a loop of interpretation, action, observation, correction, and continuation.
What this direction studies
The direction has a stable scope even as individual questions and publications change. Six areas organize the work.
Goals and planning
How agents translate an objective into a sequence of steps, account for constraints, revise plans as conditions change, and recognize when a task is underspecified or infeasible.
Tool and action selection
How agents decide whether to retrieve information, call a tool, use software, ask a question, wait, or take no action—and how they choose among tools with different capabilities, costs, permissions, and failure modes.
State and memory
How systems preserve the parts of previous work that matter: observations, decisions, unresolved questions, environment changes, tool results, commitments, and recoverable progress across longer tasks.
Verification and recovery
How agents inspect outcomes, test whether an action produced the intended state, detect failure, retry safely, reverse an action where possible, and recover without repeating the same mistake.
Autonomy, review, and stopping
How a system decides what it may do independently, what requires approval, when it should escalate, and when continuing would create more risk or cost than value.
Coordination
How tools, specialized agents, human reviewers, and changing environments fit into one decision process without obscuring responsibility or fragmenting the state of the work.
Evaluation cuts across all six areas. A useful assessment should examine not only whether the final task succeeded, but also the quality of the plan, the actions taken, policy adherence, unnecessary steps, recovery behavior, use of resources, and the amount and timing of human oversight.
Questions guiding the work now
The current agenda does not claim that these questions are solved. It identifies where this direction is being deepened.
When should an agent act, ask, or stop?
How should a system decide between taking an action, retrieving more information, asking a clarifying question, requesting approval, refusing an unsafe instruction, or ending a task that cannot be completed reliably?
How should planning work when tools are imperfect?
Agents operate with tools that may be irrelevant, unavailable, broken, expensive, inconsistent, or unable to perform the requested action. How should plans adapt without producing unnecessary calls or pretending that an impossible task remains feasible?
What must an agent remember across long work?
Which observations, decisions, dependencies, and environment changes must persist so that an agent can resume coherently without carrying forward stale assumptions, lossy summaries, or irrelevant history?
How should agents verify and repair their own work?
What checks are appropriate before and after an action? When is a retry justified? When should the system choose another route, reverse a change, or escalate rather than continue repairing indefinitely?
How should decision quality be evaluated beyond task success?
A correct final result can still come from a poor trajectory. What should be measured about planning quality, tool choice, policy adherence, recoverability, security, cost, latency, unnecessary actions, and oversight?
How should agents handle untrusted information and shared responsibility?
Tool outputs, webpages, documents, messages, and other agents may contain incorrect, conflicting, or adversarial instructions. How should a system separate data from authority while keeping responsibility visible across agents, tools, and human reviewers?
Our current working position
The present inAi position is provisional: useful agency does not come from autonomy alone. It comes from a decision process that can connect goals to actions, observe consequences, preserve relevant state, check itself, recover, and expose the right points for human or system review.
The appropriate degree of autonomy is therefore contextual rather than universal. A low-impact, reversible task may justify wider freedom. A consequential, ambiguous, or difficult-to-reverse action may require stronger constraints, clearer evidence, or approval before execution.
Research should be able to challenge this position. Evidence may show that some tasks need different architectures, that some review gates add little value, or that apparently strong recovery mechanisms fail over longer horizons. The position is a frame to test, not a conclusion to protect.
How we approach the work
This direction combines conceptual and applied research.
Literature and benchmark review
We examine published work on planning, tool use, memory, feedback, evaluation, agent–environment interaction, human oversight, and the security of tool-calling systems.
Architecture and failure analysis
We compare how agent systems represent goals, state, tools, permissions, verification, and recovery, and where those structures fail under longer, noisier, or more consequential work.
Evaluation and protocol design
We develop questions, taxonomies, and protocol frames that help separate final outcome from decision quality, trajectory quality, recoverability, and oversight.
Applied observation
Where a product, prototype, or Open Source tool creates a genuine test surface, practical behavior can expose assumptions that remain invisible in abstract discussion: unclear instructions, stale state, tool failures, review costs, and difficult recovery paths.
Every substantive output should distinguish external evidence from inAi interpretation. It should state its publication type, date, version, scope, limitations, and review status. Product observations can generate research questions, but a product is not treated as proof that a research claim is settled.
Public work
This collection is intentionally small. It lists one inAi publication today and will grow only as additional work is actually ready.
Agentic Decision Systems — from assist to autonomous execution
The October 2025 overview presents the first public framework for this direction. It focuses on plan–act–verify–repair–resume loops, autonomy gates, oversight measures, audit trails, rollback, drift, and the application of agentic decision systems to bounded operational work.
- Publication type
- Foundational overview
- Publication status
- Published
- Published
- 1 October 2025
- Data vintage
- October 2025
- Review status
- Published by inAi; not presented as externally peer-reviewed
- Included framework
- Initial ADA-1 Agentic Decision Audit
It records the direction as it stood in October 2025. The living agenda on this page can evolve without rewriting that historical publication.
Read the foundational overviewThe overview includes the initial ADA-1 Agentic Decision Audit framework.
Additional work will be added here when it is published. Depending on the work, the collection may include review papers, research notes, evaluation protocols, applied studies, or essays. Each item should be labelled by its actual type and review status.
Where this direction sits in the research program
Agentic Decision Systems is the action and control layer in inAi's research map. It connects vertically to the systems view of intelligence and horizontally to the other three research directions.
AGI as a System
Agents matter because a capable model is only one part of an intelligent system. Continued work also depends on tools, memory, state, execution, feedback, coordination, permissions, and an environment in which actions have consequences.
Limits of Intelligence
Long-running agent work exposes limits in planning, context, tool use, memory, verification, and recovery. Studying those failures helps distinguish apparent autonomy from reliable capability.
AI for Knowledge Creation
Knowledge-oriented agents must decide what evidence to gather, which source to trust, what hypothesis to test, whether a contradiction has been resolved, and when a conclusion is justified strongly enough to present.
AI and Business Operations
Operational environments give agent decisions real constraints: permissions, handoffs, policies, deadlines, cost, review, compliance-sensitive work, and consequences for people and organizations.
From research to agent-operable software—and back
Research does not serve as decoration for product claims, and products do not prove the research. The relationship is iterative.
Agentic Decision Systems can inform how agent-facing software represents tools, state, permissions, review, recovery, and continued work. Products, prototypes, and Open Source implementations can then reveal where those assumptions break: where documentation is unclear, state becomes stale, a tool returns an ambiguous result, recovery is too expensive, or human oversight arrives too late to be useful.
Products for Agents is inAi's product category for this application layer. Software for the agent era explains the broader product thesis: agents are becoming software operators and need products designed for the way they discover, call, inspect, and continue work.
Selected external research
This is not an inAi publication list and not a complete literature review. It is a compact set of external work that helps define the research landscape. The list should be reviewed as the field changes.
ReAct: Synergizing Reasoning and Acting in Language Models
A foundational reasoning-and-action framework that interleaves model reasoning with actions in an external environment.
Reflexion: Language Agents with Verbal Reinforcement Learning
Research on language-based feedback and episodic memory as a way for agents to improve decision-making across repeated attempts.
OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
An execution-based benchmark for agents operating real computer applications across open-ended tasks.
τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
A benchmark for agents that must interact with users and programmatic tools while following domain rules and policies.
AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents
A benchmark for measuring both task utility and security when tool-calling agents operate with untrusted data.
Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents
A June 2026 preprint that separates planning quality from execution and tests planning under irrelevant tools, broken tools, constraints, and infeasible tasks.
AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications
A 2026 preprint focused on memory across long agent–environment trajectories involving states, actions, observations, and tool outputs.
Research collaboration
Agentic decision systems improve through criticism, comparison, and practical testing. inAi is open to serious proposals from universities, laboratories, independent researchers, technical partners, and institutions working on planning, tool use, memory, evaluation, agent security, human–agent coordination, or agent-operable software.
Relevant collaboration may include a joint literature review, external critique of an inAi framework, evaluation design, protocol development, benchmark analysis, an applied study of a bounded agent workflow, or student research around a clearly defined question.
A useful initial proposal should explain the research question, why it fits this direction, the proposed form of collaboration, the people involved, the intended output, and any data, publication, confidentiality, or funding requirements.
From capable models to accountable action
Agentic systems will matter not only because they can produce better answers, but because they can make and carry decisions across time. The research problem is to understand what makes that process coherent: how agents plan, use tools, preserve state, verify outcomes, recover from failure, coordinate with others, and stop when continuing is no longer justified.
That is the purpose of Agentic Decision Systems at inAi.