Workflows agentiques — Maîtrise8 / 8
Multi-Tool Pipelines — Ticket to Reviewed Branch
The payoff: chain MCP, hooks, commands and subagents into one flow that takes a ticket to a reviewed branch — with you in the loop only where it counts.
Each building block is useful alone. Chained, they become a pipeline that takes a ticket from "assigned" to "reviewed branch" while you watch the decisions, not the keystrokes.
The flow
- Pull the work — an MCP server fetches the ticket (Linear, Jira, GitHub) so the agent has the real spec, not your paraphrase.
- Plan first — start in
planmode. The agent reads the codebase and proposes an approach. You approve or redirect — the one human gate that matters. - Implement with guardrails — it edits code; a
PostToolUsehook auto-formats every file; aPreToolUsehook blocks anything destructive. - Review by a specialist — a
code-reviewersubagent reads the diff in its own context and reports bugs, security issues, and missing tests. - Ship — a
/open-prcustom command writes the branch, commit, and PR body from the diff.
Wiring it as one command
Package the repeatable part as a slash command so the team runs the same pipeline:
---
description: Take a ticket to a reviewed branch
argument-hint: [ticket-id]
allowed-tools: Read, Edit, Bash(git *:*), mcp__linear__*
---
1. Fetch ticket $1 via the Linear MCP and summarize the spec.
2. Propose a plan and wait for approval.
3. Implement; rely on hooks for formatting and safety.
4. Invoke the code-reviewer subagent on the diff.
5. Open a PR with a summary of changes and the review findings.
Unattended runs
For CI or overnight batches, run headless: claude -p with --output-format json and a strict --allowedTools. The agent never blocks on a prompt — because you set the limits up front.
Série — Workflows agentiques — Maîtrise
- Partie 01L'état d'esprit du flux de travail agentiqueLa plupart des développeurs traitent l'IA comme un autocompléteur plus intelligent. Ceux qui prennent l'avance la traitent comme un coéquipier qui agit. Voici le modèle mental.
- Partie 02MCP Servers 101 — Give Your Agent Real ToolsLe Model Context Protocol est la façon dont votre agent arrête de deviner et commence à interroger votre base de données, vos tickets, votre navigateur. Voici le modèle mental et la première connexion.
- Partie 03Build Your First Custom MCP ServerLes serveurs prêts à l'emploi couvrent GitHub et Postgres. Le vrai levier, c'est le serveur que seul vous pouvez écrire — le pont vers votre propre système.
- Partie 04Hooks — Make the Agent Obey Your RulesA prompt asks the model to remember. A hook makes it happen — deterministically, every time, outside the model's control.
- Partie 05Commandes slash personnalisées en tant que flux de travail d'équipeUne commande slash personnalisée est un prompt réutilisable que vous engagez dans le dépôt — ainsi toute l'équipe exécute la même instruction de haute qualité au lieu de la retaper.
- Partie 06Subagents — Déléguer du travail qui s'adapteUn contexte géant devient lent et vague. Les subagents permettent à l'agent principal de confier du travail ciblé à des spécialistes dotés de leur propre contexte et outils — et de les exécuter en parallèle.
- Partie 07The Daily-Driver Setup — Settings, Permissions, Status LineThe difference between fighting the agent and flowing with it is twenty minutes of configuration you do once. Here is the setup.
- Partie 08Multi-Tool Pipelines — Ticket to Reviewed Branch — vous êtes iciThe payoff: chain MCP, hooks, commands and subagents into one flow that takes a ticket to a reviewed branch — with you in the loop only where it counts.