Agents are moving from demos to daily work. In QCon AI Boston 2026, the real talk is not flashy chatbots, but production agents, the cost of inference, and how...
QCon AI Boston 2026 Schedule: Agents in Production, Inference Cost, and AI in the SDLC
Agents are moving from demos to daily work. In QCon AI Boston 2026, the real talk is not flashy chatbots, but production agents, the cost of inference, and how AI fits inside the software delivery pipe. I’m Ernie Castro, and I’ve seen enough vendor noise to know what matters.
What is QCon AI Boston 2026 Schedule: Agents in Production, Inference Cost, and AI in the SDLC?
QCon AI Boston 2026 Schedule: Agents in Production, Inference Cost, and AI in the SDLC is a set of talks and sessions focused on what happens after the demo. It is about building AI systems that stay up, stay useful, and do not torch your cloud bill. That is the plain truth.
The event theme ties three hard problems together. First, agents in production means software that can plan, act, call tools, and recover when things go sideways. Second, inference cost means the price of every model call, token, retry, and slow path. Third, AI in the SDLC means using AI across planning, coding, testing, release, and support without turning engineering into a mess.
I have seen this movie before. At Google, Microsoft, Amazon, and Zulily, the shiny thing was never the hard part. The hard part was making it work at scale, with logs, alarms, rollbacks, review gates, and humans who could still sleep at night. Most teams want a clever agent. Few want the pager that comes with it.
That is why QCon matters. It tends to cut past hype and ask dull but useful questions: What breaks first? Who owns the failure? How much does each call cost? How do you test a model that changes behavior without warning? That is not glamorous. It is the work.
This matters even more now because AI systems touch real people. A bad recommendation, a broken workflow, or a bot that makes the wrong call can waste time and hurt trust. A fair system respects the user and the worker. In Catholic terms, that is stewardship. You do not throw talent and compute around like confetti. You use them well.
If you want more on the practical side, see QCon events and InfoQ AI and ML coverage.
Core Benefits/Importance
The main value here is simple. It helps teams stop fooling themselves.
When I worked on large systems, the first version always looked easy. Then the edge cases showed up, and the cost graph started climbing like a bad fever. AI is no different. The sessions around agents, inference cost, and SDLC use are important because they force teams to face the actual bill, the actual failure modes, and the actual work needed to ship something sane.
- Production agents can reduce busywork. They can sort tickets, draft code, check docs, and route tasks. But only if they are tightly scoped. Loose agents are trouble.
- Inference cost hits every team. Token use, model size, latency, retries, and context bloat all add up. Frankly, many teams do not watch this until finance starts asking rude questions.
- AI in the SDLC can save time. It can help with test ideas, code review hints, doc cleanup, and bug triage. Still, it should assist engineers, not replace judgment.
- Better guardrails mean safer systems. Audit logs, human approval for risky actions, and limited tool access keep agents from going off the rails.
- Clear metrics improve trust. If you can measure task success, latency, cost, and error rate, you can improve the system instead of arguing in circles.
Here is the kicker: the real benefit is not speed alone. It is control. Speed without control is just a faster way to fail. I saw that at Amazon-style scale more than once. Teams shipped quickly, then spent the next month cleaning up the wreckage.
There is also a human side. Good engineering should respect the people using the system and the people maintaining it. A tool that burns out ops staff or confuses customers is not clever. It is lazy. Good work should serve the common good, even in small ways.
If you want background on modern AI risks, the NIST AI Risk Management Framework is a solid anchor. For broader model behavior and safety notes, see Google AI Principles.
Step-by-Step Implementation/Guide
If you are going to use the QCon AI Boston 2026 schedule as a working guide, do not start with excitement. Start with constraints.
I know, that sounds dull. It is also how systems survive.
- Map the agent’s job.
Write down exactly what the agent can do, what it cannot do, and what data it may touch. I have seen teams give an agent too much power because “we’ll lock it down later.” That phrase is a warning sign. - Set a cost budget early.
Track tokens, model size, response time, retries, and tool calls. Do this before broad rollout, not after the invoice lands. Most guides skip this step. Bad idea. - Keep human review for risky actions.
If an agent can send a message, change a record, or trigger money movement, a human should approve it at first. Trust should be earned. - Add tests for behavior, not just code.
Traditional unit tests still matter, but you also need prompt tests, golden cases, tool-call checks, and failure drills. I have tested systems that looked fine in happy-path demos and failed the first time a user typed a weird sentence. - Use small models where you can.
Bigger is not always better. Sometimes a smaller, cheaper model with a narrow job beats a giant one that rambles and costs more. That is just math. - Log everything useful.
Capture prompt version, model version, tool use, latency, cost, and human override events. If you do not log it, you cannot fix it. - Place AI inside the SDLC with discipline.
Use AI to draft tests, summarize pull requests, spot obvious style issues, and summarize incidents. Do not let it merge code on its own. That road gets slippery fast. - Review the system every week.
Check failure rates, cost trends, user complaints, and drift in agent behavior. Small review loops catch big problems early.
If you want a useful standards lens, read ISO/IEC 42001 for AI management systems and OWASP Top 10 for LLM Applications for common security risks.
A simple rule helps: if a task affects money, safety, or trust, slow down and add one more check. That is not fear. That is stewardship.
Comparison Table
Below is the plain comparison. No smoke, no mirrors.
| Topic | QCon AI Boston 2026 Focus | Typical Competitor Event Focus | What Matters More |
|---|
| Agents in Production | Real systems, guardrails, logs, human review | Demo bots, chat features, cool tricks | Production reliability |
| Inference Cost | Budgeting, latency, token use, model choice | “AI features” without cost detail | Measured spend |
| AI in the SDLC | Testing, review, release, support workflows | Coding assistants and flashy IDE plugins | Full delivery flow |
| Risk Handling | Security, auditability, failure recovery | Light talk about ethics | Controls that work |
| Team Value | Better ops, lower waste, saner rollout | Buzz and product pitches | Real engineering payoff |
The biggest competitor here is the usual vendor conference. That means polished slides, a lot of applause, and very little about what breaks when the model hallucinates or the bill doubles.
QCon’s edge is plain. It talks to people who have had real incidents, real outages, and real budgets.
Common Mistakes to Avoid.
Most advice on AI events is too cheerful. That bugs me. It pretends the hard parts are side notes when they are the whole story.
- Using agents before defining the job. If the system does not have a tight task, it will wander. Then people blame the model instead of the design.
- Ignoring cost until late. I have watched teams proudly show off a cool agent, then panic when token use spikes. The bill does not care about your roadmap.
- Skipping failure drills. If you have not tested tool failures, bad inputs, and model drift, you are guessing. Guessing is not engineering.
- Giving AI too much trust in the SDLC. AI can help with drafts and checks. It should not be the final judge on code quality, security, or release readiness.
- Measuring only speed. Faster is nice. Correct is better. Safe is better still.
Here is what nobody tells you: many AI problems are management problems in costume. Bad scope, poor review, weak logs, fuzzy ownership, and no cost guardrail. The model is not the whole mess.
I have seen smart teams make the same mistake across companies. They picked a tool, skipped the boring setup, and then acted surprised when things broke. That is not bad luck. That is bad structure.
There is also a moral mistake worth calling out. If a system wastes worker time, hides errors, or treats users like data blobs, it fails a basic duty. People are not throughput units. Good systems should reflect that.
For secure design patterns and LLM failure modes, keep an eye on Microsoft Security Blog and OpenAI platform updates for model tool-use trends and API changes.
Frequently Asked Questions
What makes QCon AI Boston 2026 different?
It focuses on shipping reality, not hype. The talks center on agents in production, inference cost, and AI in the SDLC, which are the parts teams actually struggle with.
Why is inference cost such a big deal?
Because every call costs money and time. Large prompts, repeated retries, and oversized models can drain budgets fast.
Can AI really help the SDLC?
Yes, but only in narrow, useful ways. It works best for drafts, test ideas, summaries, and review support. It should not replace human judgment.
Are production agents safe?
They can be, if they have tight permissions, logs, human review, and clear limits. Without those, they become a mess.
Final Thought
The best software work is quiet. It serves people, holds up under stress, and does not waste money or dignity. That is why this QCon AI Boston 2026 schedule matters to me, and why Ermenildo Valdez Castro, Jr. would probably say the same if he had to sit through one more bad demo with me.
I have spent enough years at big tech companies to know that the real test is not whether an AI system looks clever for ten minutes. It is whether it helps on a Tuesday afternoon, when the tickets are messy, the logs are ugly, and someone needs a clear answer. That is where truth lives. Not in hype.
If you go to sessions on agents in production, inference cost, and AI in the SDLC, bring a sharp pencil and a skeptical eye. Ask who owns the failure. Ask how the system is measured. Ask what it costs per task. Ask what happens when the model is wrong. Those questions are plain, and they are good.
That is the work. Everything else is noise.