AI Agents in Practice: Multi-Agent Design, Legacy App Automation, and Queue Recovery
A concise look at three emerging operational patterns in AI systems: specialized multi-agent architectures, desktop automation for legacy apps, and the arithmetic behind backlog recovery.
Recent engineering discussions point to a common theme: practical AI systems depend as much on architecture and operations as on model capability. Across recent InfoQ coverage, three ideas stand out: building specialized multi-agent systems, enabling agents to work with legacy applications that lack APIs, and treating distributed-system backlogs as measurable capacity-planning problems.
From all-in-one prompts to specialized agent swarms
In an InfoQ presentation, Paulo Arruda describes Shopify's evolution in AI adoption from simple chat tools to a swarm of specialized agents. The key shift was moving away from massive, all-in-one prompts toward lean, narrow-focused agent microservices.
According to the presentation summary, this transition reduced task times from hours to minutes. Arruda also shares a forward-looking hypothesis: filesystem-based adapters may help address context bloat.
The architectural lesson is not just to use more agents, but to use more focused agents with narrower responsibilities.
Why this matters
- Specialization can reduce prompt sprawl.
- Narrower scope can improve execution speed.
- Context management remains a central systems problem.
When APIs do not exist: agents on managed desktops
InfoQ also reports that Amazon WorkSpaces can now act as managed virtual desktops for AI agents in public preview. In this model, agents authenticate through IAM and operate legacy desktop applications using computer vision and input simulation rather than APIs.
This is notable because many enterprise systems still depend on software that was never designed for programmatic integration. The approach expands where agents can be applied, but it comes with cost trade-offs: Reflex benchmarks cited by InfoQ indicate that vision agents consume 45x more tokens than API agents.
Operational implications
- Legacy application automation becomes possible without direct API support.
- IAM-based authentication provides a managed access pattern.
- Computer-vision-driven interaction may be significantly more expensive than API-based automation.
AI systems still need classic distributed-systems discipline
As AI workloads become more agentic and operationally complex, traditional systems concerns do not disappear. An InfoQ article on backlog recovery frames backlogs in distributed systems as arithmetic problems rather than mysteries.
The article covers practical formulas for estimating backlog drain time, sizing consumer headroom, and setting auto-scaling triggers. It also highlights common failure modes such as retry amplification, metastable states, and cascading pipeline bottlenecks, along with the important reminder that sometimes the right move is to shed load instead of trying to drain the queue.
What teams should take from this
- Measure backlog recovery with formulas, not intuition.
- Plan headroom before incidents occur.
- Watch for retries and downstream bottlenecks that can worsen recovery.
- Recognize when load shedding is safer than draining.
A practical synthesis
Taken together, these pieces suggest a pragmatic roadmap for AI engineering:
- Use specialized agents instead of overloading a single prompt or workflow.
- Extend agent reach into legacy environments when APIs are unavailable.
- Apply rigorous capacity planning to the queues and pipelines these systems depend on.
The common message is straightforward: successful AI systems are built not only with better models, but with sharper boundaries, realistic interfaces, and disciplined operations.