Delegate and Run Work in Parallel

More agents do not automatically produce more useful work. Two tasks can run in parallel only when their inputs are available, their owned surfaces do not conflict, and their results can be integrated through a known contract. If one task defines the schema another must use, doing both simultaneously creates rework disguised as speed.

Draw the dependency graph before delegating. Parallelize independent research, isolated components, separate test investigations, or adversarial review. Keep tightly coupled implementation sequential until the shared interface is accepted. Include integration and final verification as real work rather than assuming separate outputs will assemble themselves.

Claude Code offers several parallel patterns. A subagent is delegated work in its own context and returns a result to the main conversation. Separate sessions are independent workstreams a human can monitor. Agent teams coordinate multiple Claude Code instances through shared tasks and messaging. Agent view provides a surface for managing many sessions. Each has different setup, communication, and context behavior.

Use the lightest mechanism that fits. A read-only investigation often belongs in a subagent because its many file reads stay out of the main context. Two independent feature branches may be clearer as named worktree sessions. A coordinated team is justified only when workers genuinely need shared tasks or peer communication. Experimental or plan-specific features should be checked against the current availability documentation before becoming a required workflow.

A custom subagent does not automatically receive the main conversation history or every file already read. Claude composes a delegation message, and the worker starts from its own context. Give it the outcome, relevant paths, constraints, tools, acceptance criteria, expected return artifact, and what it must not do. If context is too important to summarize, keep the task in the main session or use an appropriate fork pattern.

Restrict tools to the job. A research subagent may need Read, Grep, and Glob but no editing. A verification agent may run tests without owning the fix. A documentation worker may edit only docs. Tool limits reduce accidental authority and make the returned evidence easier to trust.

When workers edit source, give them isolated worktrees or non-overlapping ownership. Claude Code custom subagents can set `isolation: worktree`, and separate sessions can start with `--worktree`. Isolation prevents direct overwrites but does not eliminate logical conflicts: two branches can still make incompatible assumptions about the same interface.

Assign one integration owner who understands the accepted specification and base branch. Require each lane to return a commit or artifact, changed surfaces, tests, assumptions, and unresolved risk. Integrate in dependency order, inspect each diff, resolve conflicts deliberately, and run the full verification suite after assembly. A successful lane test is not proof that the combined system works.

One valuable use of parallel context is independent verification. The agent that implemented a change has seen its own reasoning and may repeat the same assumptions when reviewing it. A fresh subagent can compare the diff with the original acceptance criteria, search for missed call sites, run adversarial examples, and attempt to disprove the completion claim.

Independence requires a different task, not just a different agent name. Give the verifier the specification, diff, proof claims, and authority to inspect, but not an instruction to confirm success. Ask for counterexamples, missing evidence, security or data risks, and exact file references. The integration owner decides which findings require repair before merge.