Interactive debugger and execution playground for AI agents. Visualize an LLM agent's plan as a live DAG, inspect every step, pause execution, steer decision branches, override node outputs, replay runs, and export an auditable Markdown trail.
Status: working concept prototype for AI-agent transparency, human-in-the-loop control, and reproducible execution. Not production-ready yet.
Most AI-agent observability tools help you inspect traces after a run. TACG explores a complementary workflow: understand and influence an agent while it is running.
- See the plan — the agent turns a task into a directed acyclic graph (DAG).
- Watch execution — nodes and edges update as steps run and the plan adapts.
- Stay in control — pause, stop, retry, override outputs, or choose a decision branch.
- Recompute safely — downstream nodes are recalculated after an intervention.
- Replay any run — event-sourced history supports timeline scrubbing and graph versions.
- Audit behavior — export decisions, interventions, model provenance, and LLM usage to Markdown.
- Bring your own model — use the default provider or an OpenAI-compatible endpoint.
| User | Job to be done |
|---|---|
| AI-agent engineers | Debug plans, branching logic, prompts, and expensive steps |
| Product and UX teams | Prototype where an agent should decide and where it should ask |
| AI consultants | Demonstrate agent behavior and deliver a reviewable run artifact |
| Educators | Teach agent planning, adaptation, and human-in-the-loop workflows visually |
flowchart LR
U[User task] --> P[Planner]
P --> G[Live DAG]
G --> E[Executor]
E --> S[Supervisor]
S -->|adapt plan| G
S -->|decision gate| H[Human or auto choice]
H --> G
E --> V[Event stream]
V --> R[Replay and audit export]
A run follows a transparent loop: plan → execute → inspect → adapt or ask → continue → replay. Every structural change and user intervention becomes an event that can be folded back into the run state.
| Capability | What TACG provides |
|---|---|
| Live computation graph | React Flow DAG with node states, animated edges, versions, zoom, and minimap |
| Human-in-the-loop control | Pause/resume/stop, manual or automatic decisions, mixed branch selection |
| Node intervention | Inspect inputs and outputs, retry a node, override its result, recompute descendants |
| Adaptive planning | Supervisor can add steps or decision gates as new information appears |
| Time-travel replay | Event-sourced timeline, snapshots, graph version switching, run history |
| Audit artifacts | Markdown export with chronology, branches, interventions, model provenance, and usage |
| BYOK routing | Default provider plus OpenAI-compatible endpoints such as OpenAI, OpenRouter, Groq, Ollama, LM Studio, or vLLM |
- Bun
- A modern browser
- Optional: an API key or local OpenAI-compatible endpoint for LLM mode
git clone [https://github.com/danku13/TACG.git](https://github.com/danku13/TACG.git)
cd TACG
bun install
bun run devIn a second terminal:
cd mini-services/tacg-agent
bun --hot index.tsOpen http://localhost:3000. The web UI runs on port 3000; the agent engine uses Socket.IO on port 3003.
Start with Demo mode to explore TACG without an LLM key.
Next.js / React 19 / React Flow / Zustand
│ Socket.IO
▼
Bun agent mini-service
planner → executor → supervisor → decisions
│
▼
RunEvent[] + fold → replay / audit
Main locations:
| Area | Path |
|---|---|
| UI and graph components | src/components/tacg/ |
| Client logic and types | src/lib/tacg/ |
| Runtime state | src/store/tacg-store.ts |
| Agent engine | mini-services/tacg-agent/ |
| Product and architecture docs | docs/ |
| E2E scenarios and screenshots | scripts/ |
| Document | Purpose |
|---|---|
| Product requirements | Problem, users, scenarios, scope, and limitations |
| Functional requirements | Capability-level requirements and status |
| Architecture decisions | ADRs and technical rationale |
| Product and market hypotheses | Segments, value propositions, and validation hypotheses |
| Change log | Requirement and architecture changes |
TACG is a concept prototype, not a production agent orchestrator. It currently has no production authentication or multi-user collaboration, stores run history in local JSON, and does not yet provide a hardened persistent database or enterprise deployment model.
- Shareable runs and multi-user review
- Production persistence and authentication
- Streaming node outputs
- More native model-provider adapters
- Multi-run comparison and experiment workflows
- Integration adapters for existing agent runtimes
Issues, architecture feedback, and reproducible agent-debugging scenarios are welcome. Before opening a pull request, review the requirements and architecture decisions.
No open-source license has been selected yet. Until a license is added, the repository remains publicly viewable but all rights are reserved by default.