← Selected work
CASE 01Enterprise SaaS engagement / 2026

From stalled prototype to repeatable delivery.

An event-driven agent fleet that moves a ticket through planning, implementation, review, and human approval.

MY ROLE

AI enablement, architecture & rollout

Multi-agent systemsClaude CodeCosmosMCPJira / GitHub

01 / The challenge

Where the work began.

The client's autonomous delivery pipeline was not reliably firing. Multiple agent fleets, brittle configuration, and unclear operational ownership made it difficult to distinguish a workflow problem from an integration failure.

02 / The approach

Agentic delivery, end to end.

I traced Jira state through webhook delivery, trigger evaluation, and agent sessions, isolating stale event-source identifiers. With the pipeline running end to end, I consolidated the fleets and made the workflow portable enough for other product teams to adopt.

DELIVERY / WALKTHROUGHInteractive example
01Ticket scopedIntent + acceptance criteria
02Plan reviewedHuman approval gate
03Agent buildsIsolated implementation
04Change verifiedTests + human review
A useful workflow starts with a clear task. Explore the checkpoints.

Illustrative workflow. No agents are running and no data leaves your browser.

03 / Engineering decisions

The choices that mattered.

01

Make state explicit

Ticket-driven state transitions, durable external state, and per-agent idempotency make retries understandable and recoverable.

02

Put judgment at the gates

Two mandatory human approval points and bounded retries keep consequential decisions visible to the people responsible for the code.

03

Design for the next team

Environment-based configuration, team-scoped run state, service-account credentials, and reusable wiring references support repeatable onboarding.

04 / The result

What shipped.

  • Moved from a non-firing prototype to production rollout and team onboarding in under five weeks.
  • Restored repeated end-to-end completion, including approval gates and terminal states.
  • Consolidated two fleets and retired six superseded agent bundles.
  • Delivered reusable onboarding guides, configuration references, and live configuration drift checks.

Client identity and implementation identifiers are omitted. This case study describes my contribution to a consulting engagement.

05 / What stayed with me

The hard part of an agentic system is often the surrounding software: state, credentials, retries, ownership, and the path back when something goes wrong.
Keep exploring / 02AI Hub

Start a conversation

Talk about the work.

AI engineering, agentic systems, developer tooling, or a problem worth thinking through.

A little context goes a long way.

Opens your email app with a draft. Nothing is sent or stored by this site.