Skip to content

Panelists Question How Far AI Agents Can Take Software Delivery Into Production

An InfoQ roundtable on AI-assisted software delivery found teams are using agents for research and deployment loops but still hesitate to hand over the middle of the lifecycle. Governance, testing and cost are the main concerns.

3 min readBased on 1 source0 views
DEVOPS

TL;DR

  • Teams use AI agents for research and for the repetitive ends of delivery, but most still keep engineers in the middle of the work.
  • Panelists say trust, governance and risk analysis decide how far agents can go into production.
  • Testing, security and token costs are becoming operational concerns, not side issues.

From solo coding to shared pipelines

An InfoQ roundtable moderated by editor Renato Losio looked at how AI is changing the path from code to production. The panelists were Patrick Debois, who works with Tessl on a software factory for coding agents; Martin Reynolds, Field CTO at Harness; Garima Bajpai, a long-time enterprise engineering leader; and Gunnar Grosch, a developer advocate at Sinch.

Debois described a progression that many teams follow. It started with code completion, moved to coding agents running on a laptop, and then to shared context such as skills and better harnesses. About six or seven months before the discussion, his team moved to a central pipeline, which he called a factory. Work there starts with tickets, and agents handle much of it, so less of it happens on individual machines. He said this changed how product, engineering, sales and support work together.

Where the hand-off gets hard

Grosch said the easiest parts to hand to agents are research before coding and the deploy, test and iterate loop at the end. The middle, where code is written and changed, still needs engineers to do much of the work. In his view, the core issue is trust. As he put it, handing off the whole job is "a bit scary."

Renato Losio noted that many practitioners are experimenting and running small projects, but few run agent-driven delivery fully in production. Reynolds said trust is especially important in regulated environments. He argued that if an organization already has solid delivery governance, the same controls can apply to agents: scoped credentials, policies, approvals and an audit trail. The goal is to avoid free-roaming agents, while accepting that they may fail in different ways than people do.

Reynolds also named code review, test generation and automatic repair of failed builds and deployments as areas where AI can help. He pointed to the growing number of CVEs that AI-driven tooling surfaces in builds as a sign of rising workload.

Testing, security and cost

Bajpai said code is no longer the scarce capability in enterprises, so the bottlenecks have moved. She argued that testing has to change for probabilistic LLM-based applications. In her account, teams move from statistical thresholds toward golden data points and LLM-as-judge rubrics. She also said security must extend into runtime, adding a shift-right emphasis to the familiar shift-left approach. Her third point was cost: at the scale of a 20,000-developer organization, token spending becomes a large line item, and FinOps practices need to cover it.

Debois agreed that good engineering basics still matter. Better context, tests and observability reduce the likelihood of problems. He also described risk scoring for changes. A documentation edit and a change to authentication code should not follow the same path. Teams then weigh the likelihood and impact of a failure, as they have long done in DevOps. He added that people who say they no longer read agent-written code often run low-stakes applications, where that risk profile is acceptable.

Why it matters

For developers, the discussion points to a practical shift. Agents are moving from individual helpers to parts of a shared delivery system, and that brings process questions with them. Teams will need clear permissions, approval steps and audit logs for agent actions. They will also need test strategies that suit non-deterministic outputs, security checks at runtime, and budgets for token usage.

For engineering leaders, the message is that the hard part is less about generating code and more about governing the path to production. Risk tiers that match changes to their potential impact may help decide where agents can act with less supervision and where engineers should stay closely involved. The panel did not present a single answer, and most participants described partial adoption rather than full automation.

Sources

This article was written by Blozhik from the reporting below. Read the originals for full detail.

InPresentation: AI-Assisted Software Delivery: From Code to ProductionInfoQ
Was this useful?

Keep reading

Discussion · 0

Sign in to join the discussion. Liking does not need an account.Sign inRegister
Loading comments…
blozhikDeveloper news from across the web, explained. Every story links to its original sources.