QCon AI Boston: Production AI Moves beyond Prompts to Platforms, Harnesses, and Evals

QCon AI Boston 2026, covered by InfoQ writer Tatiana Fesenko, centered on the operational hard parts of putting AI agents into production rather than on prompt craft alone. The through-line was that production AI now depends on platforms, harnesses, and evals—the systems around the model—more than on one-off prompt tuning.

Speakers framed the work as infrastructure: better context management so agents keep the right state across steps, a security harness that constrains what an agent can call and access, and a full engineering model that treats agents like other production software. That model covers how you design, ship, observe, and evaluate agent behavior instead of treating each run as an isolated chat session.

For engineers and builders, the shift is concrete. You stop optimizing only the prompt and start owning the control plane: context windows and memory strategy, tool and permission boundaries, and eval suites that catch regressions before users do. Teams that already run services will recognize the same demands—reliability, least privilege, and measurable quality—applied to agent workflows.

The conference framing also maps onto a broader market move: vendors and platforms are competing less on “best prompt” and more on production-ready agent stacks—orchestration, policy, observability, and evaluation. Organizations choosing tools will weigh how well a stack supports harness-style security and continuous evals, not only raw model quality.

What to watch next is whether your own agent work has those three layers in place. Audit context management, define the harness around tools and data access, and treat evals as a release gate. If any of those are still ad hoc, the QCon AI Boston message is that they are now first-class production concerns, not nice-to-haves after the demo.