Agently

Agently is a durable execution platform for long-running, multi-agent workflows. The API and web app form the control plane; Temporal, the Python reasoner, and the Node pieces worker form the execution plane.

Architecture

Next.js web -> Go API -> Postgres
                           |
                    queued runs
                           v
               Python reasoner dispatcher
                           |
                 Temporal workflow history
                           |
               one activity per DAG node
                  /          |          \
              Gemini     Browserbase   pieces worker

• apps/api: workflow planner, CRUD, schedules, triggers, logs, credentials, and the web API.
• apps/reasoner: the only run executor. It walks DAGs in deterministic Temporal workflow code and persists each node result as an activity checkpoint.
• apps/pieces-worker: executes pieces.* actions/triggers and stores durable trigger state.
• apps/web: live UI backed only by API data.
• packages/nodes: hand-written and generated integration catalogs.
• packages/db/migrations: the Postgres schema and upgrades.

The retired native Go execution worker and the runs.engine selector have been removed. Migration 0013_drop_retired_execution_columns.sql also removes its lease columns.

Strict failure model

Agently does not manufacture successful-looking work:

• Planner router/map/reduce failures return the provider error.
• Missing Gemini or Moonshot (Kimi) credentials prevent planning.
• Missing or stale piece embeddings prevent API startup.
• Missing Browserbase, Langfuse, SMTP, database, or integration credentials fail startup or the selected node.
• An unavailable pieces worker fails piece nodes.
• The web app does not synthesize costs, logs, browser sessions, or activity data.

Models

The planner uses PLANNER_MAP_MODEL (default gemini-3.1-flash-lite) for routing and map selection, and PLANNER_MODEL for graph creation and repair (code default kimi-k3 via the Moonshot API; .env.example uses gemini-3.5-flash so a single Gemini key runs everything). The provider is inferred from the model name, and any tier can run locally as ollama/<model> (e.g. ollama/qwen3:8b): no keys or quotas, slower on a laptop. PLANNER_EMBED_MODEL provides catalog recall — ollama/nomic-embed-text locally, or gemini-embedding-001 / text-embedding-3-small. The reasoner uses REASONER_MODEL and REASONER_SYNTHESIS_MODEL (default gemini-3.5-flash-lite). Gemini 429s are retried after the server-suggested delay, so free-tier keys work, just slower.

Local development

cp .env.example .env
# Fill every required credential.
pnpm install
pnpm agently

Open the web app at http://localhost:3000 and Temporal UI at http://localhost:8080. See RUNBOOK.md for setup, migrations, service restart, and verification commands, and its "Deploying" section for running the stack in containers behind an access token and password.