The production framework
for AI agents.
CrewAI and LangGraph frame how your agent thinks. agentFast frames everything above that — and stays agnostic about which one runs underneath. This is what a run looks like once it does:
| planning | complexity | medium — budget raised to 30 iterations | — |
| memory | recall | 2 facts about cust_alice | — |
| guardrail | pii:input | 1 email redacted before the model saw it | — |
| llm | claude-sonnet-4-5 | 1 tool call | $0.0031 |
| tool | kb_search | ok · 3 hits · 6ms | — |
| llm | claude-sonnet-4-5 | 1 tool call | $0.0028 |
| tool | order_lookup | ok · ord_1001 · 4ms | — |
| hitl | approval_requested | refund_request — run paused, durably | — |
Not a developer? Start here — no code on that page. Every capability below also opens with a plain-language summary before the technical detail.
Get started
From a clean machine to an agent answering a real ticket.What an agent is, what goes wrong, and why any of this exists. No code.
One command to a running agent, playground and dashboard.
Install into your own project, with the extras per SDK.
Scaffold an agent, point it at your data, gate a risky tool.
What changes and what doesn't across the five adapters.
Build
Make it yours: your tools, your behaviour, your infrastructure.The production layer
The capabilities agentFast frames — identical on every SDK.The four layers, the adapter seam, and the hole at L1.
Checkpoint per iteration; resume from the exact step after a kill.
Pause on risky tools, survive a restart, execute exactly once.
Watch a run live over SSE — tokens, tools, approvals, re-attach.
One error vocabulary, MCP servers by config, agents as MCP servers.
Long-term facts, session buffer, episodic log, and compaction.
PII redacted before the model sees it, injection defence, limits.
Step trees, cost and latency per step, OTEL and LangSmith export.
A judge that can't be contaminated, and a CI gate that fails closed.