Your first agent

agentfast new scaffolds a project from a template: a config file, a knowledge-base folder, seeded eval cases, and a Compose file. You then replace the knowledge base with your own and adjust which tools need a human.

Scaffold

agentfast new support --sdk langgraph
cd support-agent

Templates available today:

| Template | SDKs | |---|---| | support | langgraph, vanilla, crewai, claude_agent_sdk, openai_agents_sdk | | research | vanilla | | code | vanilla |

You get:

support-agent/
├── agentfast.yaml       # model, budget, guardrails, which tools are risky
├── kb/                  # your documents go here
├── evals/cases.yaml     # seeded cases — the quality gate
└── docker-compose.yml

Point it at your data

Drop .md or .txt files into kb/. They're ingested and embedded on startup — no separate index step:

cp ~/company/refund-policy.md kb/
cp ~/company/shipping.md kb/
docker compose up

The agent's kb_search tool now answers from your documents, and every answer carries a grounding score so you can tell whether it was actually supported by what it retrieved.

Decide what needs a human

This is the part worth thinking about. In agentfast.yaml:

tool_overrides:
  refund_request:
    risk: high          # pause the run and wait for a human
    hitl_mode: suspend
  send_email:
    risk: high
    hitl_mode: defer    # queue the call, let the run continue

The two modes answer different questions:

• suspend — the agent cannot sensibly continue without knowing the answer. The run pauses, checkpoints, and survives a restart. Use it for anything that moves money or is hard to undo.
• defer — the call is queued for approval but the run carries on. Use it when the agent has other useful work to do and the action can happen later.

Anything not listed is low-risk and executes normally.

Run the quality gate

agentfast eval

Runs the cases in evals/cases.yaml, scores each against your thresholds, prints a per-metric breakdown, and exits non-zero if anything is below bar — so it works as a CI step without any further wiring. See Evals for how the judge avoids grading its own homework.

Then