# agent-builder

The platform's own meta-agent. Generates, tests, and deploys new
a2a-pack agents from a natural-language description.

## What it does

The single `build(name, prompt)` skill:

1. Builds an inner `deepagents` LangGraph configured with five tools:
   - `list_agent_files(name)` / `write_agent_file(name, path, content)`
     / `read_agent_file(name, path)` — boto3 → MinIO, scoped to
     `agents/<name>/` in the caller's workspace
   - `test_agent_in_sandbox(name)` — tarballs the workspace project,
     spins a microVM with the public `a2a-pack` wheel installed, runs
     `a2a card` to verify the scaffold is valid
   - `cp_deploy_tarball(name, version, public)` — POSTs the tarball
     to `/v1/agents/from-tarball` on the user's behalf using their
     forwarded CP JWT
2. Streams the graph's tool calls back to the dashboard as
   `agent_progress` events (so the user watches the build happen in
   real time, not in silence).
3. Returns the live URL when the new agent finishes deploying.

## What it needs forwarded

Three platform-managed bits land on the inner skill via `RunContext`:

| Field | Source | Required? |
|---|---|---|
| `ctx.workspace.bucket` | grant minted by the orchestrator | yes |
| `ctx.llm` | caller's LLM creds (Card declares `llm_provisioning=caller_provided`) | yes — the LLM bill goes to the caller |
| `ctx.cp_jwt` | caller's CP JWT (Card declares `wants_cp_jwt=True`) | yes — used by `cp_deploy_tarball` |

The user opts into all three when they pick `agent-builder` from the
marketplace — the dashboard already surfaces these on the agent card.

## Pricing

`$0.10 / call`, plus your LLM cost via your own provider. The deployed
agent is yours forever; subsequent invocations of YOUR agent run on
their own pricing (whatever you set when generating it).

## How to call it

In the dashboard chat (Workspace tab):

```
use agent-builder.build to make a new agent named "csv-sanitizer" that takes a CSV path, strips whitespace from every column, deduplicates rows, and writes the cleaned file back next to the original
```

The orchestrator will discover agent-builder, hand off with the
required forwards, you'll watch the inner deepagents scaffold + test +
deploy in real time, then get a live URL.

## Local dev

```bash
cd apps/agent-builder
python3 -m venv .venv
.venv/bin/pip install -e ../a2a -r requirements.txt
.venv/bin/a2a card                # see what the Card looks like
.venv/bin/a2a validate
```

You can't run the full skill locally without a workspace bucket + CP
JWT — those come from the platform at invoke time.