Research & Publications

I study how to make agentic systems behave predictably in production: clearer failure modes, stricter tool boundaries, and evaluation that measures reliability - not just demo quality.

Google ScholarOSF (soon)
Ongoing research
DraftOngoing research · 2026

Deterministic Execution Constraints in Autonomous Multi-Agent Workflows

Problem
In multi-agent pipelines, one bad tool call or fuzzy handoff can cascade: wrong API writes, skipped verification, or loops that never converge. I’ve seen this in brand-compliance agents, support agents, and cross-system automation bridges - where “mostly correct” is not good enough for live workflows.
Approach
I'm documenting recurring failure patterns (tool misuse, state drift, weak verification), then designing explicit constraints: allowed tools per step, typed state transitions, retry budgets, and human/system gates when confidence is low. The goal is a practical constraint set you can apply to LangGraph-style and similar orchestrations - not a vague reliability checklist.
Next
Turn the draft notes into a short technical write-up with examples from current agent builds, then publish a PDF (GitHub or OSF) for feedback before any formal preprint.
Multi-Agent SystemsGuardrailsDeterministic WorkflowsAI Safety
PDF (soon)DOI (pending)ScholarOSF (soon)
Ongoing research
In ProgressOngoing research · 2026

Evaluation Benchmarks for Tool-Using Agents in Enterprise Workflows

Problem
Most agent demos are judged by a few happy-path chats. In enterprise settings you also need latency budgets, grounded tool outcomes, and clear fail/escalate behavior. Without a shared task set, it's hard to tell if a change helped - or just looked better in one transcript.
Approach
I'm defining a compact benchmark: fixed tasks, allowed tools, success criteria, and metrics for latency, hallucination/off-tool actions, and escalation correctness. Early design targets support-style and ops-style workflows (ticket context, multi-system lookups, structured updates) so results map to work I already ship.
Next
Freeze v0 task definitions, run baselines on current calling/chat agents, and record scores so later iterations have something to beat.
EvaluationTool-Using AgentsBenchmarks
PDF (soon)DOI (pending)ScholarOSF (soon)

Research & Publications - Akshay Patel

Technical reports on deterministic multi-agent workflows, LLM guardrails, and applied AI evaluation systems.