Orchestration

Set up evo in a repo by building its benchmark and running the first experiment

Run structured experiment loops that try variants and keep the top candidates

Summarize evo run results and frontier candidates without running any benchmarks

Protocol for evo optimization subagents dispatched from the optimize loop

Hand code analysis and refactors to the Codex CLI from inside your session

Set up and run external coding agents such as Codex over the Agent Client Protocol

Spawn and manage background subagents that run work in parallel

Split a big job across parallel subagents that run unattended and report back

Manage AI coding agents on parallel tasks with high-level direction and automated review

Run multi-agent workflows defined in YAML steps with the TAKT engine

Show live status of every agent, their tmux panes, tasks and unread inboxes

Coordinate independent subagents in parallel and spawn workers for multi-part tasks

Split two or more independent tasks across parallel subagents when delegation is cheaper

Run plan tasks through independent subagents with review between each task

Trace an agent session into a ledger of the choices it made

Run progressive experiments to map how tunable parameters affect a goal

Run Claude Code as an independent subagent for review or delegated tasks

Get a second opinion from the Codex CLI on reviews and delegated changes

Carve risky parts out of a task and run the safe remainder in a fresh subagent

Check for tunnel vision by fanning out fresh reads and reporting where they diverge

Review one plan or file through independent lenses and compare where they agree

Run repeated build, QA and rework cycles for long autonomous coding projects

Cold-read an artifact in a fresh session to check it stands on its own

Run autonomous AI research projects with fast experiment loops and a steering outer loop

Evaluate proposals for feasibility, engineering payoff and risk

List, inspect, stop and stream logs from Bernstein orchestrator agents

Show failed tasks, stalled agents and budget warnings from Bernstein

Review and approve or reject pending Bernstein tasks and plans before they run

Show Bernstein spend by model and agent, with budget status and projections

Create a task in the Bernstein orchestrator and queue it for an agent

Break a complex goal into staged Bernstein plans with dependencies

Show pass rates, success rates and completion times per model in Bernstein

Run a multi-agent goal in parallel git worktrees and verify the merged output

Check Bernstein status: active agents, task progress, spend and alerts

Break goals into tasks and queue them on a task server for agents

Coordinate between agent teams, resolve conflicts and decide when to pivot

Design agent loops with deterministic verifiers, stop conditions and context limits

Coordinate Deep Agents with subagents, task planning and human approval steps

Fan out many independent items to parallel subagents and combine the results

Set up a new domain loop in a file-based knowledge base and run one test pass

Delegate tasks to Codex, Aider or other CLI agents in detached tmux sessions

Delegate coding tasks to the Blackbox CLI, which compares results from several models

Delegate feature builds, refactors and PR reviews to a separate coding agent CLI

Delegate feature builds, refactors and batch issue fixes to the Codex CLI

Spawn independent long-running agent processes for tasks that run on their own

Delegate feature work, refactors and PR reviews to the OpenCode CLI

Run plans task by task with fresh subagents and a two-stage review

Offer basic care check-ins to neighbors in a simulated town

Follow up on recurring health needs and medication routines for villagers