Code is no longer the bottleneck. Everything around it is.
Atlassian measured that writing code is about 16% of engineering work, and that AI raises merged-PR throughput by around 10% even when almost 100% of PRs are AI-assisted.
This means complexity moved from coding to everywhere else.
Map where work waits (requirements, review, QA, cassette re-recording). Assign relevant KPIs and avoid getting blinded by pure code-writing efficiency increase.
Context is king. Not the code.
The complexity moved "one layer of abstraction higher", into the context, knowledge and skills a company feeds to its agents.
Agents grounded in Atlassian's teamwork knowledge graph gave 44% better answers and used 48% fewer tokens than baseline semantic search. This is especially true on relational questions like: "who owns this, what changed yesterday, what depends on this".
Treat the KB, the registry and the skills repo as production assets with the same precision and accuracy standard as code (as they have a positive compounding effect over time).
Skills proved to be a relatively stable primitive. Prompts on the other hand need continual improvement.
Skills “have maintained relative stability [...] for the last 12 months” while models and harnesses kept changing.
Track skill usage. Make the campfire "show-and-tell" sessions not only around solutions but around skills as well.
As cost per unit of work becomes critical, "tokenomics" becomes a thing.
Vercel and Dropbox run a routing layer that picks the model per task. Dropbox started a "tokenomics" practice because spend rose 28x in a year.
It's goal is to protect the gross margins and R&D efficiency.
The goal is not to limit spend. The goal is to know whether token spend goes to the right problems. Dropbox tracks LLM cost per PR and per Jira ticket. Think about ways to implement this.
Use events to trigger repetitive,
well-scoped, clearly reviewable work first.
automatically
saved
Atlassian's Jira Coding Agent now resolves 51% of potential security vulnerabilities automatically.
It has cut issue cycle time from 11.5 to 6.7 days, and saved about 50 engineering days a week. Changes labeled as low-risk go straight to the agent, high-impact changes get summarized to a human first.
Pick two repetitive, scoped and reviewable workflows and wire each to a trigger. See examples:
Sentry error → Jira ticket, drafted by an agent and checked by a human.
Cassette refresh on prompt or model changes.
In order for the approval queue not to become the bottleneck name the approver and set a deadline (SLA) for each HITL flow.
Governance is about transparency.
Atlassian puts the agent session and its activity log on the Jira work item, so the team sees what the agent was asked, what it did and what changed as a consequence.
Have written boundaries as integral parts of agent’s ID and log agent actions where the team already looks (like Jira tickets).
Review is the new bottleneck.
Atlassian's AI Reviewer checks every PR against standards defined once and against the Jira requirements. This reduced PR's cycle time by 45%.
Dropbox saw median PR size double from ~40 to ~70 lines and "change confidence" has fallen as engineers can read the code but now trust it less.
Write down once the standards you care about, such as shared prompt blocks, naming, security rules and test coverage. Let an AI reviewer check every PR against them and against the ticket’s requirements. Treat prompts like code.
Ultimately, the only outcomes that count,
are the ones for the end-customers.
100% developers use AI in their work regularly, with 6 self-reported hours saved per week, and a 40% gap between perceived and actual productivity (as actual productivity is smaller).
An innovation ratio up only 1% against 28x spend.
The reason behind these numbers is that: the new capacity is going into migrations, upgrades and quality gates that sat in backlogs for years. Still their control metric is a changelog that lists only customer-visible changes.
Make sure your work is covered by proper Analytics strategy (consult your local Product Leader for it) and track end-customer value.
Deliberately protect human-to-human collaboration, because agents pull people toward talking with them instead of real humans.
Across industry data, cross-team collaboration and deep work fell for the first time: “people are very comfortable talking to an agent [...]”. Dropbox saw this and had their deep work went up 11 points by making a deliberate effort sponsored by the Chief People Officer. Atlassian found only 48% of its team felt AI actually improved collaboration.
Agree in which phases people must talk to each other before an agent gets involved, and protect that time.
Dropbox is a useful example: it holds architecture and design reviews with senior engineers before code is written.
Feed the factory from real signals, and treat business-side input as the scarce resource.
Vercel feeds sales-call transcripts, support tickets and social signals into its backlog as prioritized context.
Atlassian captures work from Slack threads with the reasoning attached, and a triage agent fills fields and finds duplicated work streams before a human looks.
Link the transcript excerpt to each Jira item, so the next person or agent reads the original discussion, not a summary.