The agent platform with observability built in
Build agents you can watch think.
Assivy gives every agent tools, subagents, and knowledge — and records every turn as a trace with spans, metrics, and logs, so when an answer is wrong you can see exactly why.
[hook:beforeTurn] fetching thread context…
[tool:get_order] order #4821 synced
[log] 3 tools · 1.2k tokens · 4.1s
✓ done — trace 8f3a…c1Everything a turn needs
One platform from prompt to proof: configure the agent, extend it with code, and inspect what happened when it runs.
Agents with real settings
Identity, instructions, and a model list with a default — users can switch models mid-chat or stay pinned. Temperature included, reasoning models handled.
Learn moreFunction tools in TypeScript or Python
Write tools that run sandboxed, log with levels, and surface their output in the trace. What a tool said is always one click from the turn that called it.
Learn moreSubagents, two ways
Derive a narrowed helper inline from the parent, or mount a stored agent as a tool. Give the helper the cheaper model — the platform enforces the downgrade.
Learn moreTurn hooks
Run code before the turn starts: fetch context, ground the prompt, log what you did. Hooks share the same logger the tools use.
Learn moreTracing, metrics, logs
Every turn is a trace with spans, token and cost rollups, and structured logs — one filter bar across all three, so the agent you picked follows you.
Learn moreKnowledge and skills
Folders of sources feed retrieval, skills inject into turns, memories persist across them. The agent answers from your content, grounded.
Learn moreRun your first agent today
The quickstart takes you from a fresh checkout to a chatting agent with its first trace — no slides, just the product.