Multi-agent release management for a global engineering organisation

How a team of AI agents answers release-status questions from thousands of Jira issues, with policy checks before every action.

Programme managers at a global semiconductor engineering organisation assembled release status by hand from thousands of Jira issues and several internal tools. Now they ask in plain English and get a cited answer in minutes.

The challenge

Release information was spread across Jira projects, linked merges and internal tracking tools. Every status meeting started with hours of manual querying and spreadsheet work, and the answers were out of date as soon as they were shared.

What we built

A multi-agent system with distinct planner, executor and display roles. The planner breaks a question into sub-tasks, executor agents query Jira and internal tools through Model Context Protocol (MCP) servers, and a display agent turns the results into a short, cited answer.

  1. Programme managerAsks a release question in plain English
  2. Planner agentSplits it into sub-tasks and picks a proven procedure from memory
  3. Policy checkOPA rules confirm the scope before any tool call
  4. Executor agentsQuery Jira and internal tools through MCP servers
  5. Display agentReturns a cited answer and drafts any status note

Procedural memory lets the system reuse workflows that have already worked, so common questions take the same reliable path every time.

How we made it safe to run

  • Policy as code. Every action is checked against OPA rules before it runs; write actions need explicit approval.
  • Golden-dataset evaluation. Each release is tested for functional match, tool-call parameters and the path the agents took.
  • Cost and latency profiling. Token use and response time are tracked per workflow, which exposed and removed silent retries in tool-calling clients.

Outcome

Release questions are answered in natural language in minutes instead of hours of manual reporting, and every answer can be traced back to the issues it came from.

Industry
Semiconductors and high-tech
Stack
Google ADK, MCP, LiteLLM, Redis, Azure OpenAI and open-weight models, OPA

Bring us the workflow that eats your team's week.

In a 30-minute call with a senior architect, you'll learn whether an agent can take it on, what it would cost to build and run, and how we'd prove it works.

Book a discovery call