Quickstart
One command to a running agent, playground and dashboard.
The fastest way to understand agentFast is to watch a run pause for a human, kill the process, and watch it resume. That takes one command.
Run the stack
Postgres with pgvector, Redis, the agent API, the playground and the dashboard all come up together.
git clone https://github.com/abhichat85/agentfast.git
cd agentfast
docker compose up
localhost:3400 — talk to the SupportAgent and watch it work in real time.
Then the dashboard at localhost:3401 for the full step tree, cost and latency per step, and the approval queue. Both read the same API on localhost:8321.
Paste this into the playground:
I'd like a refund for order ord_1001 — the item arrived damaged.
You'll watch it size up the task, search the knowledge base, look up the order — and then stop. Refunds are configured as high-risk, so the run pauses and waits for you.
The default model is scripted-support, a deterministic stand-in that exercises the real harness
— real tools, real checkpoints, real approval gates — without calling a provider. Everything you
see is the actual production layer, not a mock of it.
Prove it's durable
This is the part worth doing by hand, because it's the property everything else is built on.
With the run paused, kill the API container:
docker compose kill api
The run is gone from memory. Bring it back:
docker compose up api
Open the dashboard. The run is still there, still paused, still holding the exact tool call it
stopped on. Approve it, and the agent continues from that step — it does not replay the knowledge
base search, and it does not issue two refunds.
This isn't a demo script. It's tests/test_flagship_kill_resume.py,
a test that runs against real Postgres with a real SIGKILL across two OS processes — and again
on the vanilla loop in test_flagship_vanilla.py.
Use a real model
Set a key and name a model. Everything else stays the same:
export ANTHROPIC_API_KEY=sk-ant-…
AGENTFAST_MODEL=claude-sonnet-4-5 docker compose up
AGENTFAST_MODEL takes any LangChain model id, so openai:gpt-4o and
anthropic:claude-haiku-4-5 work too.