AgentFlow
Deterministic orchestration for AI coding workflows.
AgentFlow
Deterministic orchestration for AI coding workflows.
AgentFlow turns specs into auditable, cost-aware development runs. The pipeline combines local investigation, git worktrees, external AI coding agents, validation commands you define, and reproducible reports so each run leaves a traceable record instead of an opaque chat log.
agentflow work "develop billing-v2"Core ideas
Local-first investigation runs before cloud model calls. AgentFlow scans the repository with bounded grep and filesystem traversal, then narrows context so you do not ship entire trees to remote APIs by default.
Cost and token estimation surfaces expected spend before execution. Heuristic counters and configurable pricing sheets give order-of-magnitude guardrails; budgets can block or confirm overruns before agents run.
Git worktree isolation keeps each task on its own branch and working tree under .agentflow/worktrees/, reducing collisions when several features move in parallel.
Agent routing maps pipeline step classes to the agent profile you configured—local Ollama for triage, cloud CLIs for heavy implementation—based on strategy and policy in YAML.
Validation before trust means AgentFlow runs your commands (go test, linters, custom scripts), not agent self-assessment, before a step is treated as done.
Reproducible reports write markdown and JSON under .agentflow/runs/<run-id>/ so you can audit what ran, what it cost, and what validation reported.
Get started
Follow these pages in order when onboarding a repository: