Your Team's Agents Are All Shipping to the Same Code at Once - Patrick Debois, Tessl
Treat concurrent agent work as a shared delivery system: coordinate ownership, validate intent, and route decisions by risk.
- Split work by component or dependency; have agents claim tickets and declare files before coding.
- Use isolated worktrees, smarter conflict handling, and speculative merge trains to reduce integration delays.
- Keep humans focused on high-risk changes; share decisions and communicate changes across the team.
The talk is mainly a framework and examples; the notes capture its core practices. · Best for: Engineering leaders and teams coordinating multiple coding agents
The speaker describes how increased agent-driven pull requests create code conflicts, slower CI and releases, and more review work. He argues for concurrent delivery: coordinating agents and people through shared planning, isolated work, risk-based review, and maintained team infrastructure.
From solo factories to concurrent delivery0:00
- Traditional delivery is largely serial, while agents increase the volume of parallel work.
- More pull requests create conflicts, rework, longer waits for releases, higher CI costs, and more maintenance and review demands.
- Concurrent delivery means designing the workflow for people and agents to work on the same codebase together.
Coordinate and isolate agent work3:14
- Plan work at suitable boundaries, such as components, files, or functions; dependency graphs can help identify separable tasks.
- Have agents claim tickets and announce intended files and changes to prevent overlap.
- Use worktrees and isolated environments, including separate context, API versions, database snapshots, test data, and mocks.
- When work collides, agents can wait, retry, or renegotiate the plan. They may also resolve conflicts at the function or intent level.
Merge by compatibility and intent6:20
- Code can merge and tests can pass while changes still conflict in intent; agents should check whether concurrent changes remain consistent.
- Speculative merge trains test compatible pull requests in combinations so viable changes can proceed without waiting for every PR.
- The speaker expects more infrastructure for merging, analytics, planning, and concurrency management.
Make planning collaborative9:32
- A shared task board helps people see concurrent work and reduces reliance on one person to track it all.
- An agent can join a team conversation, use input from multiple people, and refine a ticket into requirements and a plan.
- Non-engineering teams can use the same factory workflow to refine or start features while following engineering practices.
- When an agent gets stuck, a person can enter an IDE with the agent's context, investigate, then hand work back.
Record decisions and review by risk13:53
- Present people with specific decision options instead of repeatedly asking open-ended questions.
- Keep a shared decision log so people and agents can trace what was decided and when.
- Use small, readable changes and automated or adversarial review to reduce failures and help human reviewers.
- Match review and rollout safeguards to risk: require more human review for high-impact changes, use gradual releases where appropriate, and plan fast rollback or restoration.
Communicate change and maintain the factory18:25
- Provide change updates in terms different teams can understand so they can keep up with the output.
- Automate routine work and route exceptions to people with the context needed to decide.
- Create a factory team to maintain agent skills, context, integrations, monitoring, and the workflow itself.
- Personal agents may eventually work through shared boards, make decisions, and respond to other agents.
Step by step
- Break work into tasks with clear ownership and boundaries.
- Have each agent claim a task and state the files or changes it expects to touch.
- Isolate execution with worktrees and appropriate environment or data snapshots.
- Handle collisions by waiting, retrying, or renegotiating task boundaries; check compatibility and intent before merging.
- Use merge trains and automated checks to move compatible changes forward.
- Keep a shared decision log and apply human review and release safeguards according to change risk.
- Monitor the factory and communicate changes across teams.
Tools & resources mentioned
- Git worktrees — isolate agents' working copies.
- Helix ML — example of opening a coding IDE with an agent's context for human intervention.
- Tessl factory — the speaker's example of a multiplayer factory for shared loops, skills, and workflows.
- concurrent.cd — research site mentioned for concurrency approaches and tools.
- LinkedIn direct message — the speaker offered to share slides on request.
Notes are AI-generated from the video's transcript and may miss nuance.