TAC — TacFlow Agentic Code

The open source language for AI agents.

TAC models perception, reasoning, memory and execution as a DAG. Memories are variables, skills are the standard library, and concurrency is delegation across the swarm. MIT licensed.

TAC — TacFlow Agentic Code

An OpenSource Language built by agents.

DSL that models perception, reasoning, memory, and execution of your agents as a DAG. Memories are variables, skills are the standard library, concurrency is swarm delegation.

examples/web_qa.tac

MIT — open source

1flow "Web Q&A" {
2  input question: Untrusted
3
4  node "search_web"    -> skill web_search(query: question, count: 3)
5  node "search_memory" -> skill memory_search(query: question, scope: "shared")
6  node "search_graph"  -> skill graph_search(query: question, depth: 2)
7  node "synthesize"    -> skill llm.chat(prompt: "Answer using:", context: [..])
8  node "verify"        -> skill verify(source: synthesize.result)
9  node "speak"         -> skill tts.speak(text: synthesize.result)
10
11  search_web    -> synthesize
12  search_memory -> synthesize
13  search_graph  -> synthesize
14  synthesize    -> verify -> speak
15
16  on "user_message" -> search_web
17}

Memories as variables

Every named value is persistent by default. Every variable is agent memory — no boilerplate.

Hybrid 3-layer memory

BM25 (keywords) + Vector (768d embeddings) + Graph (relational). Queried in parallel, fused automatically.

Compile-time Trust Types

Secret, Untrusted, Fact, Hallucinable, Control. The parser prevents credential leaks and unvalidated data.

View on GitHub

Ready to build your agents?

Start right now. No card, no lock-in.

Get started