Agents make more custom automation economically viable, while software increasingly needs to be designed for agents as its users.
2
The main engineering bottleneck is shifting from writing code to designing systems, managing context, setting boundaries, and deciding what agents should do.
3
OpenClaw shows both the appeal and the risks of personal agents, especially when they can access private data, untrusted content, and communication tools.
Summary
The first day of AI Engineer Europe presents agents as a new application layer and as a new way to build software. Malte Ubl argues that agents make previously uneconomic automation practical, while APIs and infrastructure must become easier for agents to use. Ryan Lopopolo describes a workflow where humans design systems and delegate implementation to parallel coding agents. OpenClaw contributors show how that approach works in open source, personal automation, container deployment, security hardening, and enterprise orchestration. Other talks cover local models and Hugging Face tooling, AI employees in Slack, token-maxing incentives at large companies, and software design practices for AI-generated code. Matt Pocock argues that domain modeling, test-driven development, clear interfaces, and maintainable architecture matter more when code production is cheap. Sunil Pai closes with code-executing agents that can use large APIs through a small, capability-controlled interface.
Malte Ubl argues that agents expand the amount of software worth building. Traditional software could not economically encode every business process, but agents can handle work that previously required too many hardcoded rules. He suggests starting with narrow tasks such as compressing research before a human decision, surfacing information already scattered across issue trackers and messages, and removing repetitive support work. At Vercel, an agent handles sales inquiries, routes support requests, researches companies, and reviews abuse reports before a human makes the final decision. Ubl says an in-house support agent achieved a 90% deflection rate, allowing support staff to spend more time on difficult cases.
Software must be built for agents as well as people
Ubl says agents are becoming software users, which changes how teams should design products and infrastructure. He reports that more than 60% of Vercel.com page views in the previous seven days came from AI agents, with usage moving from dashboard clicks toward APIs and CLIs. He asks engineers to treat command-line and programmatic access as first-class interfaces. Agent applications also need sandboxes and a separation between the harness and the code it runs. His broader point is that the application layer can remain stable while models change underneath it, leaving room for engineers outside the major model labs to create new products.
Coding agents make systems thinking and delegation more important
Ryan Lopopolo says he spent nine months building software exclusively through agents and banned his team from using editors directly. He describes code as abundant and says the scarce resources are human time, human and model attention, and context windows. Engineers therefore need to design systems that agents can understand, delegate work across many parallel sessions, and make requirements legible through documentation, architecture decision records, tests, lints, and review agents. He gives a concrete example: a bespoke lint can require every network fetch to include retries and timeouts. A source-code test can also enforce a file-size limit of 350 lines so the codebase stays easier for agents to understand.
OpenClaw's growth creates a security workload that automation cannot fully solve
Peter Steinberger describes OpenClaw as five months old and growing at an extraordinary rate, with around 30,000 commits and nearly 2,000 contributors at the time of the talk. That growth brought a flood of security reports. He cites 1,142 advisories, around 16.6 per day, including 99 marked critical, with about 60% closed. Steinberger says many reports are generated by agents, but maintainers still need to read them because rushed automated fixes can break the product. He recommends treating a personal agent differently from a team agent, enabling sandboxing in group chats, and limiting the data available to shared agents. The risk is highest when an agent has private data, untrusted content, and communication abilities at once.
Personal agents become useful through small, inspectable steps
Radek Sienkiewicz explains how OpenClaw gradually became part of his daily life. He started with one chat channel and one simple workflow, then added access to email, calendars, files, operating-system automation, and an Obsidian vault containing about 3,000 notes. The agent indexes and backs up his data while he sleeps, prepares morning summaries, filters urgent messages, drafts email replies, and maintains search indexes. His advice is to start with one recurring pain, build trust incrementally, and keep the system inspectable through Markdown files. He warns that bad memory compounds as the knowledge base grows, while long brittle automations and weak boundaries create maintenance problems.
Containers provide a practical boundary for agent deployment
Sally Ann O'Malley makes the case for running agents in containers. She uses Podman and describes portability across laptops, architectures, Kubernetes, and OpenShift, along with volumes for state backup and recovery. Podman secrets keep API keys outside ordinary environment values, while OpenClaw secret references add another layer of indirection. She imagines company-approved agent images containing MCP servers, skills, authentication, and access to internal drives, which can then be adapted for individual employees. Her setup uses local containers, Kubernetes, OpenShift, SSH sandboxes, and optional OpenTelemetry tooling. The aim is a repeatable environment that can be moved from local development into managed infrastructure.
AI employees need shared context and carefully scoped access
Fryderyk Wiatrowski presents Viktor, a Slack-based AI employee with access to about 3,000 integrations. Unlike a personal agent, one company connection can provide shared access for a team, although permissions must be scoped so private email or executive context does not leak into unrelated channels. Wiatrowski says Slack makes long-running work feel more natural because people already expect coworkers to take time to respond. The design challenge is turning DMs, threads, edits, deletions, reactions, and channel messages into coherent agent context. He also says personality affects adoption. Users preferred Claude Opus in one comparison because Viktor felt more natural and less mechanical.
Software fundamentals control the quality of AI-generated code
Matt Pocock rejects the idea that teams can repeatedly compile specifications into code without maintaining the code itself. He says each regeneration made his codebase worse because software tends toward entropy when people focus only on local changes. His practical response is to use a shared design concept, a ubiquitous language, test-driven development, and deep modules with simple interfaces. A 'grill me' skill interviews the developer until the design is understood, while a ubiquitous-language skill records domain terms for both people and agents. TDD forces smaller steps and faster feedback. Pocock's recommended division of labor is to design the interfaces and delegate the implementation, while preserving clear boundaries that humans can test and review.
Code execution can replace large collections of JSON tool calls
Sunil Pai describes Code Mode, where an agent writes JavaScript and runs it against a controlled environment instead of making many individual JSON tool calls. Cloudflare's API has about 2,600 endpoints, which would require roughly 1.2 million tokens if every endpoint became a tool. A search-and-execute interface reduces that to about 1,000 tokens in his example. Code also gives the model loops, state, sequencing, parallel execution, and type checking. Pai stresses that the execution environment must begin with no capabilities and receive access explicitly. His preferred design uses fast V8 isolates, controlled network access, and detailed observability so teams can inspect what an agent did and why.
"The scarce resources in this world that we see today are three things. Human time, human and model attention and model context window."1:11:39
Who should watch
You are deciding where agents can automate a business process while keeping a human decision-maker in the loop.
Your coding agents produce large amounts of code, but your team lacks durable architecture rules, tests, or review systems.
You want to run personal or enterprise agents with access to private data and need practical boundaries around containers, permissions, and untrusted content.
