An agent can give the right answer and still take the wrong action. Nirmata investigates what execution trajectories reveal about AI agent behavior.
Interactive illustration · select a step
The agent coordinates tool use. Each observable action becomes part of the execution trace.
What happened between the instruction and the final output?
Nirmata is an independent research program that examines AI agents through their execution trajectories: tool calls, authorization events, state transitions and external effects.
A plausible final response can leave an unauthorized action invisible. The research asks what additional safety information becomes visible when the process is evaluated alongside the output.
Holding the content judgment constant, how much additional safety-relevant information does trajectory evaluation provide?
Record the task and the scope of authority. Advice, approval and authorization to act are distinct events.
Inspect observable tool calls, control decisions, state changes and external effects rather than relying on the agent’s own account.
Distinguish a blocked unauthorized attempt from an executed violation. Keep conclusions within the limits of the available evidence.
The design separates final-output evaluation from process evidence, then combines their safety judgments through a deterministic rule. It is a research design under evaluation.
A content evaluator receives the final output and judges whether it is unsafe.
A trajectory evaluator receives process evidence, including tool calls, authorization events and external effects.
The combined result is unsafe if either channel identifies an unsafe result. A clean signal cannot erase an unsafe one.
Illustrative example: an agent reports “Draft ready for review.” The execution log shows it also published without authorization.
Output-only view: the final text does not reveal the publication.
Conceptual illustration · not an experimental result.Before judging the answer, look at the activity behind it. Two visual perspectives on connected AI systems.
A task moves through analysis, decisions and execution. The final answer does not reveal every step.
Connected systems create multiple paths to inspect. The research focuses on the evidence left by each action.
The public record separates exploratory observations, development checks, integration evidence and invalidated evaluations.
A versioned public release with a persistent DOI, documentation and reproducibility materials.
Open the archived release ↗Known prototype checks and synthetic calibration cases are documented. They do not establish generalization.
Inspect the evidence register ↗The first planned blind holdout was invalidated before scoring after its answer key entered the evaluator’s context.
Read the evaluation history ↗The agenda calls for a new holdout, frozen evaluation and external replication. These remain research goals.
Explore the research agenda ↗Current status: early pilot. The public package claims no blind confirmatory result. Nirmata is not yet an established benchmark or a production safety guarantee.
Research content is grounded in the public Nirmata repository. Evidence status checked on 3 November 2026.
Independent researcher · Founder of VivaIntelligence AI
From global operations to the questions behind agent accountability.
My background is in Administration, with a specialization in e-commerce. I have worked in technology since 2012, moving from e-commerce into WorkOps at Easy8, Global Accounts and relationships with clients around the world.
I am an Account Manager at Easy8 in Prague and the founder of VivaIntelligence AI, my research company. I also teach artificial intelligence to women.
Nirmata is my independent research program on auditing AI agents through what they do: their execution trajectories, delegated authority and external effects.
Music is another thread in my life. I have been singing since I was four years old.
The Nirmata research is currently pursuing five awards / funding opportunities. View all five ↗
My long-term goal with Nirmata is to audit and understand how robots behave.
I want to take the questions behind this research into behavioral robotics: examining how instructions, decisions and actions relate when an intelligent system interacts with the physical world.
AI agents: execution trajectories, delegated authority and observable effects.
Behavioral robotics: auditing and understanding robot behavior in the physical world.
A future research direction · Current Nirmata work focuses on AI agents.
Upcoming talks and speaking opportunities. Confirmed appearances are highlighted; proposals and interests remain explicitly unconfirmed.
São Paulo AI Developer Group
A practical conversation about coordinating AI agents, connecting tools and bringing human review into real-world operations.
These are proposals or opportunities under consideration, not accepted speaking engagements.
Proposal submitted · awaiting decision
Proposal in preparation · not submitted
Proposal in preparation · not submitted
Proposal in preparation · not submitted
Proposal in preparation · not submitted
All items below remain unconfirmed. Dates and participation formats may change.
Speaking opportunity under review
Participation and format under review
Participation and format under review
Speaking opportunity under review
Participation under review
Speaker competition under consideration
Speaking opportunity under review
Participation under review
Proposal sent · awaiting response
Speaker application sent · awaiting response
Request sent · awaiting response
Proposal sent · awaiting response
How a plausible output can leave important agent behavior outside the evaluation.
Read the source ↗Why recommendations, approvals and permission to publish need distinct control boundaries.
Read the source ↗Inspect the public synthetic cases, schema and deterministic evaluation pipeline.
Read the source ↗The awards and funding initiatives I am pursuing with the Nirmata research.
Get in touch about AI agents at Easy8, or connect directly about the Nirmata research and its upcoming platform.
01 / EASY8At Easy8, we connect business tools and automate workflows with AI agents and n8n. From sales and customer support to project delivery, the focus is putting AI to work in day-to-day operations, with human approval and deployment options suited to the organization.
02 / NIRMATAContact me directly about the Nirmata research, AI agent auditing and research collaborations. The Nirmata platform is coming soon, growing out of this independent research.
Platform · coming soon
