A small Laravel chat application that demonstrates the twdnhfr/laravel-deepagents package — a deep-agent harness for the Laravel AI SDK with planning, sub-agents, persistent memory, human-in-the-loop approval and automatic context management.

It's a single-page chat that shows the package's key features working together in a real request/response flow.

• Multi-turn conversations — the agent run state is persisted in the session and continued on each message (DeepAgent::continue()).
• Tools — four demo tools the model can call: get_weather, roll_dice, delete_records and fetch_report (see app/Tools).
• Planning — withTodos() gives the model the write_todos tool; the sidebar renders its plan live from RunState->todos as steps move through pending → in progress → completed.
• Sub-agents — a report-analyst registered with subAgent(): the parent delegates report analysis via the task tool, the sub-agent works in its own isolated context window (inheriting the parent's backend) and only its final answer returns to the conversation.
• Persistent memory — memory('memory/profile.md') loads a profile from the backend into the system prompt at the start of every conversation; the agent updates it via write_artifact. Reload the page (fresh conversation) and ask "What do you know about me?" — the history is gone, the memory isn't.
• Human-in-the-loop with all three decisions — the destructive delete_records tool is gated with requireApproval(); the run suspends and the UI shows the pending call. Approve runs it as requested, edit the arguments inline before approving (RunState::edit() — the call runs with your corrected input), or reject with an optional reason (RunState::reject()) that goes back to the model as the tool's result, so it reacts in-conversation instead of the turn being dropped.
• Persistent artifacts & offloading — large tool outputs are clipped with offloadLargeToolResults() and stored as run-scoped artifacts via the DatabaseBackend, so the model can read the full text back on demand with read_artifact.
• Context compaction — summarize() (with a deliberately low budget here) compacts older history into a summary once it grows too large; the UI marks the moment with a 🗜️ line, and the conversation keeps its context.
• Tool trace — each reply shows which tools ran with their arguments and results.

The wiring lives almost entirely in routes/web.php, which is heavily commented as the reference for how to use the package.

• PHP 8.3+
• Composer
• Bun (or npm — adjust the commands accordingly)
• An OpenRouter API key

git clone https://github.com/twdnhfr/deepagents-chat.git
cd deepagents-chat
make install        # .env, dependencies, app key, SQLite DB, migrations

Then add your OpenRouter key to .env:

OPENROUTER_API_KEY=sk-or-...
# OPENROUTER_MODEL=openai/gpt-5.4-nano   # optional, this is the default

Start the dev server (Laravel + Vite):

make dev

and open the URL it prints (e.g. http://localhost:8000).

The demo uses SQLite out of the box for zero configuration. To use MySQL or Postgres instead, change the DB_* values in .env.

cp .env.example .env
composer install
bun install
bun run build
php artisan key:generate
touch database/database.sqlite
php artisan migrate
php artisan serve   # in one terminal
bun run dev         # in another

make test           # or: php artisan test

The MIT License (MIT). See LICENSE.