Workshops: Agentic Engineering Practices
Multi-Agent Software Engineering: Roles, Parallel Work, Merging
Running several coding agents at once sounds like multiplying a team's output. In practice it multiplies coordination problems just as quickly: conflicting edits, duplicated work, rising costs. This tool-agnostic workshop gives experienced developers the patterns that make multi-agent workflows pay off, and the criteria to recognise when a single, well-briefed agent is the better choice.
Who is this workshop for?
- Experienced developers who already work with a coding agent every day
- Tech leads designing how their team works with agents
- Platform and DevEx teams building agent pipelines or internal orchestration
- Architects evaluating multi-agent tooling for the engineering organisation
- AI engineers prototyping their own agent orchestration
What you will take away
- Decide with clear criteria when multiple agents pay off and when coordination costs outweigh the gain.
- Design role-based workflows with planner, implementer, tester and reviewer agents and explicit hand-off artefacts.
- Run agents in parallel in isolated workspaces such as git worktrees, containers or cloud sandboxes, with clear file ownership.
- Apply a merge strategy for parallel branches: integration order, conflict resolution and an integration test gate.
- Set budgets, stop criteria and logging so multi-agent runs stay affordable and traceable.
Agenda
Module 1
One agent or many?
- Failure modes: conflicting edits, duplicated work, context drift, multiplied cost
- A decision matrix based on task size, coupling and verifiability
- Exercise: classifying tasks from your own backlog
Module 2
Roles and hand-offs
- Planner: specification and task breakdown as a file, not a chat message
- Implementer and tester with separate instructions
- A reviewer agent using a different model or tool for a genuine second opinion
Module 3
Parallel execution
- Isolation with git worktrees, dev containers or cloud agents
- Slicing work by module ownership to avoid collisions
- Exercise: three agents on three independent tickets
Module 4
Merging and integration
- Integration order and rebase strategy
- Conflict resolution with a human in the loop
- Integration tests as the gate before anything reaches main
Module 5
Orchestration options and economics
- Built-in options: Claude Code subagents, OpenAI Codex cloud tasks, Cursor background agents, the GitHub Copilot coding agent
- Custom orchestration via SDKs, for example the Claude Agent SDK
- Token budgets, time limits and run logs
When not to use multiple agents
More agents mean more context to prepare, more output to review and more chances for conflict. In our experience, these situations are usually better served by one agent with a good plan:
- Tightly coupled changes to the same files or the same data model
- Exploratory work where the goal is still unclear
- Small tasks that one agent finishes in a single session
- Teams whose review capacity is already the bottleneck
- Codebases without reliable tests to verify parallel results
The bottleneck moves to review
With parallel agents, the limiting factor is rarely code generation but human attention. So we also work on review practice: small pull requests, clear descriptions written by the agent, reviewer agents as a pre-filter, and the rule that nobody approves more parallel work than they can genuinely understand.
Tool-agnostic on purpose
The patterns work whether your company uses Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Gemini CLI or a custom set-up. We demonstrate them with the tools you have approved and compare how each implements subagents, background tasks and isolation, so your workflows survive the next tool change.
Frequently asked questions
Which tool is best for multi-agent work?
It depends on your environment. Terminal agents with subagents suit local orchestration, cloud agents suit asynchronous parallel tickets and SDKs suit custom pipelines. We compare the options against your use cases rather than declaring a winner.
Do more agents mean faster delivery?
Only if the tasks are truly independent and your team can review the output. Otherwise you mainly get more conflicts and more review load. The workshop helps you find the point where parallelism stops paying off.
Do we need special infrastructure?
Not to get started. Git worktrees and local containers are enough for the exercises. For larger set-ups we discuss cloud sandboxes, dedicated runners and the security requirements that come with them.
Workshop
Get in touch
How many agents does your next feature really need?
Tell us roughly how many people will join and pick dates that work – we will get back to you within 24 hours with a proposal for content, format and schedule.
Prefer to talk first?
15 minutes, free and without obligation – with our AI Trainers and AI Code Engineers.
Book a call for free (opens in new tab) +49 761 519 751 50 (24/7)Request a training / workshop
Workshop: Multi-Agent Software Engineering: Roles, Parallel Work, Merging · 1 day
More workshops
- Workshops: Agentic Engineering PracticesBug Triage & Backlog Reduction with AI Agents: Workshop
- Workshops: Agentic Engineering PracticesMVP in a Week with Coding Agents: 2-Day Greenfield Workshop
- Workshops: Agentic Engineering PracticesAgentic Coding Dojo: Katas for Regular Team Practice
- Workshops: Agentic Engineering PracticesTDD with Coding Agents: Hands-On Red-Green-Refactor Training
- Workshops: Agentic Engineering PracticesLarge-Scale Refactoring & Codemods with AI Agents: Workshop
- Workshops: Agentic Coding by Tech StackAgentic Coding with Vue & Nuxt: Hands-On Workshop
