Authority is not identity
Knowing which agent made a request does not prove that a person authorised the action or that the request stayed within scope.
Siri (Asiri) Dalugoda
I build and govern software systems where technical decisions have operational consequences.
I am a technology executive and researcher based in Auckland, New Zealand. My career began in engineering and moved outward into architecture, delivery leadership and research. I write about engineering systems, autonomous agents, AI security and the software people actually have to operate.
The thread connecting those areas is control. I am interested in how intent becomes execution, how authority moves through a system, how quality is independently verified and how organisations retain accountability as automation becomes more capable.
This site is an independent publication. It carries the essays, the research notes and the Agentic Sprint essay. The methodology that essay introduces, thirteen documents and nine working templates, is published in full at theagenticsprint.com.
From engineering to delivery systems
My career has moved outward from writing software to designing the technical and organisational systems that allow software to be delivered safely.
Learning how systems fail
I began as a software engineer and worked through enterprise software and integration environments at Virtusa and IBM. That foundation still shapes how I approach architecture: start with the operating constraints, make failure visible and keep the system understandable to the next person who has to change it.
Moving from code to organisational systems
Later roles at EROAD, ProjectManager, FLEETCOR and Futureverse widened the problem from building software to organising the conditions around it: teams, roadmaps, architecture dependencies, quality, release decisions and the flow of work across complex programmes. Most delivery failures I have watched were not engineering failures. They were decisions nobody owned.
Governed autonomy and the evidence it needs
My current work is about systems that act: how intent becomes execution, how authority travels with a delegated action, and what evidence has to survive that journey. The writing and research here follow the same question into agentic delivery, AI security and the operational software where those ideas meet real work.
Security when software can act
My agentic-security work starts from a simple concern: authentication tells a system who is present, but not necessarily who authorised a delegated action. As work moves from a person to an orchestrator, sub-agents and tools, authority can become detached from the original intent.
Within Helixar's research programme, I designed Human Delegation Provenance (HDP) and authored the 2026 research paper and IETF Internet-Draft. HDP records delegated scope as an append-only chain of signed hops that can be verified offline. It is provenance evidence, not an enforcement system by itself. A downstream service must still decide what to permit. HDP-P extends the research towards physical agents and actions whose consequences may be difficult or impossible to reverse.
Evidence must survive delegation
When one agent delegates to another, the chain should remain tamper-evident and independently verifiable at the point of action.
Enforcement belongs at the boundary
Provenance supplies evidence. The receiving application still needs an explicit gate that allows, rejects or escalates the requested action.
Automation receives no ambient authority
Agents and background processes should operate inside the same role, data and approval boundaries that constrain human users.
The same boundaries in operational software
The same philosophy applies in a far less speculative setting. In the operational job-management software I have built, roles and permissions are enforced at the data and API boundaries, financial totals are calculated on the server, and automated tools remain constrained by the authority of the user and organisation they represent. A client interface, or an AI assistant, should never be able to grant itself more power by changing a request.
That approach is informed by years of working around financial systems, enterprise platforms and operational teams: security is strongest when it is part of the system's structure, not a promise made by the interface.
Open security engineering
These repositories are published openly and can be inspected and run by anyone. Together they cover delegated authority, MCP security, release integrity and practical tools for teams working with agents. They are published through Helixar Limited, an interest declared at the foot of this page.
- HDPAn open protocol and implementation for preserving human delegation provenance across chains of AI agents.
- SentinelA security scanner for MCP configurations, live endpoints and container deployments.
- ReleaseGuardAn artifact policy engine for inspecting, hardening, attesting and verifying release outputs.
- MCP security checklistA practical security baseline for teams deploying Model Context Protocol systems.
- UnpinchedA focused detection utility for PinchTab activity in AI-adjacent developer environments.
- Helixar MCPAn MCP connector that brings Helixar security tools into agent and developer workflows.
Open-source contributions
Contributions made upstream to projects owned by other maintainers. A contribution does not imply ownership, partnership or endorsement.
docs(performance): add Step 4 (verify with keep-or-revert)
A merged contribution to Addy Osmani's agent-skills repository that documents a keep-or-revert step for measured performance work.
Authored by asiridalugoda and merged upstream on 26 July 2026.
Proposal: Selectively Preserve Tiny, Scoped Repository-Memory Capsules
An open agent-skills proposal about preserving small, scoped repository-memory capsules across agent sessions.
Opened by federicobartoli, not by Siri. Siri contributed implementation evidence from Autonomous Loop and design critique in the discussion. The issue remained open as of 29 August 2026.
Add Gemma 4 + HDP agentic security example (FunctionGemma/)
An HDP agentic-security playbook for Gemma 4 contributed to Google's Gemma Cookbook. The merged example lives in the repository at apps/Gemma_4_HDP_Agentic_Security and ships a notebook, a README and HDP middleware.
Authored by asiridalugoda from the Helixar-AI fork of the cookbook and merged upstream on 16 April 2026. A merged contribution does not imply Google partnership or endorsement.
What I keep writing about
- Agentic systemsAutonomous agents, orchestration, memory and bounded execution.
- EngineeringArchitecture, quality, tooling, reliability and software delivery.
- AI governance & securityAuthority, provenance, evidence, controls and accountability.
- Delivery & organisationsOperating models, teams, decision-making and technical leadership.
- Notes & experimentsShort observations, prototypes, questions and things learned while building.