Portfolio

Durable Workflow

2026 – now · Design lead

Demos · Docs · Workflows that wait for people (Medium) · Video: human tasks · Video: an agent that follows the policy · module-ballerina-workflow

Long-running business processes in WSO2 Integrator, on durable execution with Temporal.

• ballerina/workflow: workflows are Ballerina functions. State is persisted, so a workflow can wait for events and people, retry steps and resume after a restart.
• Human tasks: role-based assignment, a task inbox, and decisions recorded in the workflow history.
• Durable agents: AI agents on the same runtime; see Durable Agents.
• Low-code: workflow, task and agent nodes, their forms, and a workflow overview in the WSO2 Integrator designer, all editing the same source.
• Operations: the Workflow Management API to inspect, retry and recover instances; traces, metrics and a task-decision audit trail in the Integration Control Plane.

I designed the module API and its runtime model, implemented the core, and review the designer and Control Plane designs.

Durable Agents

2026 – now · Design lead · part of Durable Workflow

Demo: an agent that follows the policy · Video · Docs

AI agents that run on the workflow runtime, so an agent's work is as durable as a workflow's.

• Durable runs: the conversation and tool results are persisted; after a restart the agent resumes instead of starting over.
• Human in the loop: the agent can ask a person and wait, as a human task, for as long as the answer takes.
• Approval gates: actions with side effects wait for a person to approve them.
• Durable agentic workflows: a workflow uses agents as steps, next to events, human tasks and other steps.
• Low-code: agent nodes and forms in the WSO2 Integrator designer.

Agent-Flavored Markdown (AFM)

2025 – now · Originator

Specification · wso2/agent-flavored-markdown · Original prototype (hasithaa/afm) · Paper (PDF)

AFM is an open specification for defining AI agents in Markdown. YAML front matter holds the configuration: model, tools, interfaces and limits. The Markdown body holds the agent's role and instructions in natural language. Keeping the two apart makes an agent readable by domain experts and portable across frameworks.

I started AFM in 2025 as a prototype. My colleague Maryam Ziyad Mohamed took it forward, and it is now developed at WSO2 with a reference implementation. We described it in a paper at the AgentCraft workshop at ACM IUI 2026:

Maryam Ziyad Mohamed, Hasitha Aravinda, Rania Khalaf. Agent-Flavored Markdown: Natural Language Specifications for Framework-Agnostic Agent Development. AgentCraft 2026: Workshop on Agentic AI Systems Development, ACM IUI 2026, Paphos, Cyprus. CEUR-WS Vol-4266.

Copy BibTeX

@inproceedings{mohamed2026afm,
  title={Agent-Flavored Markdown: Natural Language Specifications for Framework-Agnostic Agent Development},
  author={{Ziyad Mohamed}, Maryam and Aravinda, Hasitha and Khalaf, Rania},
  booktitle={Joint Proceedings of the ACM IUI 2026 Workshops co-located with the 31st Annual ACM Conference on Intelligent User Interfaces (IUI 2026)},
  editor={Tommasel, Antonela and Kahr, Patricia and Dinakar, Karthik and others},
  series={CEUR Workshop Proceedings},
  volume={4266},
  publisher={CEUR-WS.org},
  issn={1613-0073},
  year={2026},
  url={https://ceur-ws.org/Vol-4266/paper03.pdf}
}

Ballerina Language & Compiler

2016 – now

ballerina-lang · ballerina.io · Language specification · Talk: Programming Languages and Compilers · Ballerina identifiers · The setup-ballerina GitHub Action

Compiler and runtime work since the early, DSL-like versions of Ballerina, and the language specification through its iterations, re-implementations and design verifications. I still work on the language and its tooling.