Research
We study agents before we trust them.
Research at Quadra Lanka is practical. Every question comes from a system we are building or plan to build, and every answer ends up in how we design the next one.
Areas
What we are working on.
Five questions that decide whether an agent can be trusted with real work in Sri Lanka and beyond.
Multi-agent coordination
How several agents divide a task, hand work between them and avoid repeating or undoing each other. Classical multi-agent systems research meets today's language-model agents.
Reliability and evaluation
Measuring whether an agent does the job the same way every time, how it fails, and what tests should pass before it touches real work.
Agentic commerce and payments
What a store, a catalogue and a payment flow need when the buyer is an agent acting for a person, and how trust and consent are recorded.
Agents in regulated finance
Human checkpoints, audit trails and explainable decisions for agents working near credit, collections and customer money.
Local-language agents
Agents that understand and answer in Sinhala and Tamil, in text and voice, including mixed-language and romanised writing.
How we work
From the bench to the build.
We read the classical multi-agent systems literature alongside the new work on language-model agents, because many of today’s problems were named decades ago: coordination, trust, negotiation and who is accountable when an agent acts.
We then test ideas on small, real tasks with measurable results. What survives goes into our toolkits and client builds. What fails gets written up too, because a clear negative result saves the next team a month.
We are open to working with universities, students and other teams on these questions. Write to us if you are working on something close.
Research notes
Written up.
Evaluating an agent: tests that should pass before real work
A practical set of tests every AI agent should pass before it touches customers, money or records, and how to build them from your own cases.
Negotiating agents: an old research field with new relevance
Researchers studied software that bargains long before chat models. As agents start buying and selling, their ideas matter again.
Coordination: how several agents avoid tripping over each other
When several AI agents share one job, they duplicate work, clash and loop. Here are the coordination patterns that prevent it.
Working on something close?
We are glad to compare notes with researchers, students and teams working on agents, commerce or local-language AI.