Concurrent coding agents
Firebender offers simple primitives to do this:| Isolation | When to use | |
|---|---|---|
| Tabs | None (same directory) | Lightweight and good for changes in completely different files, or for read-only operations like exploring the codebase |
| Background agents | Full (worktrees) | Independent features/PRs that can have conflicts and you don’t want affecting your cwd. Auto heals from conflicts from the configured base |
| Heavy mode | Managed by AI | Complex tasks where you want AI to decide how to parallelize, and cost is not a concern |
Zone of productivity.
Common pitfalls
Parallelization overhead
Problem: Conflicts, higher error rates, and context-switching costs can eat into productivity gains. Solution:- Use Background agents for isolated changes that auto-heal merge conflicts from base
- Use Heavy mode to let AI orchestrate parallelization for complex tasks
- Prefer smaller, narrowly-scoped tasks: they have the highest merge rate and lowest context-switching cost
- Custom agents: create sub-agents scoped to specific operational tasks or subset of changes to improve accuracy
Custom agents scoped to specific tasks (refactors, PR review, doc sync) dramatically improve accuracy. Create your own with
/agent or see Custom agents.Staying in flow
Problem: Managing AI agents can feel like being an engineering manager: frequent context switches, waiting on slow responses, and forgetting your original intent (“doorway effect”). Solution:- Planning mode: AI researches approaches and asks clarifying questions before coding
- GLM 4.6 and GPT-5.2: fast agentic models for quick iteration
- Commands: quickly insert pre-written prompts or task descriptions you use often
Use
/help to quickly create a command based on a previous chat. See Commands.