Keeping Humans Accountable as AI Agents Take On the SDLC

# What this is about Ming Wu, head of engineering for Dev AI at Atlassian, discussed how teams can scale AI agents across the software development lifecycle while keeping people accountable. The conversation focuses on two linked capabilities: governance and agent loops, plus the practical choices for organizing agent-driven work.

# Governed agent loops in plain terms

Put together, the approach lets agents carry out a sequence of steps — pick up an issue, perform the work, open a pull request — while the system records what happened and applies rules to prevent unsafe or noncompliant actions. The human obligations do not disappear: people remain accountable for the results.

# Why use existing work-tracking systems Jira is an example of the initial interface for agentic workflows because it already records work across teams and organizational boards. Using a system of record reduces fragmentation: the system logs agent activity, ties changes back to work items, and provides an auditable trail for reviewers and stakeholders.

That choice treats agent activity as another form of work that must be visible, assigned, and reviewed, rather than as ad-hoc or isolated automation.

# Shared context layer: a practical solution to coordination When multiple agents contribute to the same body of work, coordination becomes an engineering problem. A shared context layer gives agents a consistent understanding of team priorities, organizational requirements, and business intent. It standardizes inputs so agents act on the same facts instead of diverging interpretations.

This layer helps reduce rework and conflicting changes, and it supports traceability by anchoring agent actions to the documented context for a task.

# What remains hard: verification and measuring value Customers are asking two measurable questions: how to scale these workflows, and how to demonstrate real improvements in delivery speed. Completing more automated tasks is necessary but not sufficient. Organizations must measure how automation affects verification, testing, approvals and overall delivery outcomes.

As AI raises the volume of machine-generated changes, verification becomes a bottleneck. Teams need evidence that agent-produced work meets quality, security and compliance standards, and they need metrics that relate automated output to delivery speed and business impact.

# Practical implications for teams

• Treat agent output as work that must be traced, reviewed, and approved through existing processes.
• Add governance controls that enforce constraints and provide corrective mechanisms when agents deviate.
• Invest in a shared context layer to reduce cross-agent friction and align agent decisions with business priorities.
• Define measurement criteria beyond task counts: measure delivery cycle time, verification throughput, and defects or rework attributable to agent actions.

# Bottom line AI agents can automate more of the SDLC, but responsibility stays with people. The technical challenge is making agent work visible, coordinated and verifiable. Using established work-tracking systems, governed agent loops, and a shared context layer gives organizations a practical path to adopt agentic workflows while preserving accountability and demonstrating whether automation delivers real improvements.