Can a clean agent finish your quickstart? A pristine Daytona sandbox, an
autonomous agent, and Braintrust-scored checkpoints. No pretending — every pass is backed
by command output.
verified — Followed the Fireworks Serverless Quickstart exactly as written. The OpenAI Python SDK was already installed in the sandbox. I created the quickstart.py script verbatim from the docs (using the OpenAI SDK approach with base_url="https://api.fireworks.ai/inference/v1", model="accounts/fireworks/models/glm-5p3-flash", and the FIREWORKS_API_KEY environment variable). Running the script returned a chat completion: "¡Hola! That's 'hello' in Spanish..." confirming the API call to the Fireworks-hosted model succeeded. No deviations were required.
friction report
Documentation fetch returns cluttered HTML navigation instead of clean quickstart text: Serverless Quickstart - Fireworks AI Docs Documentation Index Fetch the complete documentation index at: /llms.txt Use this file to discover all available pages before exploring further.
Authentication setup is implicit and not shown in commands: api_key=os.environ.get("FIREWORKS_API_KEY")
Quickstart requires installing the OpenAI SDK instead of a Fireworks-specific package: cd /home/daytona && pip3 install openai
Model identifier uses a long, namespaced path that is easy to mistype: model="accounts/fireworks/models/glm-5p3-flash"
API endpoint requires overriding the default OpenAI base URL: base_url="https://api.fireworks.ai/inference/v1"
stuck — I successfully set up the CopilotKit runtime endpoint, which demonstrably works: `curl http://localhost:3000/api/copilotkit/info` returns HTTP 200 with valid JSON showing `{"version":"1.73.3","agents":{"default":{"name":"default","className":"BuiltInAgent"...}}}`. However, the Next.js frontend page is stuck in compilation (`○ Compiling / ...`) and never responds, so the full app does not serve. I had to deviate from the docs by changing the model string from `"openai:gpt-5.4-mini"` (which does not exist) to `"openai:gpt-4o-mini"`, and no real OpenAI API key was present in the sandbox environment.
friction report
Missing OpenAI API key prerequisite: OPENAI_API_KEY is set:
Length: 0
Docs model string typo: I had to deviate from the docs by changing the model string from "openai:gpt-5.4-mini" (which does not exist) to "openai:gpt-4o-mini"
Build process killed: npm run build ... Killed
Next.js catch-all route shell escaping: mkdir -p /home/daytona/my-copilot-app/app/api/copilotkit/\[\[...slug\]\] and GET /api/copilotkit 404 in 5.9s
Frontend dev server compilation hang: ○ Compiling / ... and curl to localhost:3000/ returned 000 after timeout
stuck — I successfully installed the Daytona Python SDK and created sandboxes via the SDK—creation demonstrably works, returning sandbox IDs like `0a46a97d-0dd9-45e3-bac1-92963fe874f6` with state `STARTED`. However, executing code inside the sandbox is blocked by a persistent network-level failure: both `sandbox.process.exec()` and `sandbox.process.code_run()` fail with `ConnectionResetError: [Errno 104] Connection reset by peer` when connecting to `https://proxy.app.daytona.io/toolbox`. I confirmed this is not an SDK bug by reproducing the same TLS handshake failure with `curl` to that endpoint. I also briefly hit a "Total CPU limit exceeded" error after creating multiple test sandboxes, which I resolved by deleting them via the REST API, but the proxy connection issue remains the blocking problem.
friction report
Outbound TLS to proxy.app.daytona.io is blocked: curl: (35) Recv failure: Connection reset by peer in connection to proxy.app.daytona.io:443
The daytona binary inside the sandbox is a server, not a management CLI: listen tcp :22222: bind: address already in use
SDK signals sandbox ready but proxy connection still fails: Sandbox is ready! ... Error: ('Connection aborted.', ConnectionResetError(104, 'Connection reset by peer'))
Required DAYTONA_API_URL env var is missing but defaults silently: DAYTONA_API_URL: Required api URL ... DAYTONA_API_URL: NOT SET
Sandbox creations accumulate and require manual cleanup: Total sandboxes: 10
stuck — I fetched the ElevenLabs quickstart, installed the SDK (`elevenlabs` and `python-dotenv`), and created `example.py` exactly as documented. The API key was present in the environment (51 characters, `sk_` prefix). However, running the quickstart's exact code with `model_id="eleven_v3"` consistently failed with HTTP 401 "Invalid API key" from the ElevenLabs API. I verified this was not an env-loading issue by testing the key directly with curl and hardcoding it in scripts—all returned the same 401 for `eleven_v3`. Notably, when I tested alternative models (`eleven_turbo_v2_5` and `eleven_multilingual_v2`) with the same key and voice, the API calls succeeded immediately (output showed `SUCCESS with eleven_turbo_v2_5` and `SUCCESS with eleven_multilingual_v2`), indicating the key is valid but the `eleven_v3` model specified in the quickstart is either not accessible with this key tier or the model ID in the docs is incorrect. I did not reach the point of saving or playing audio because the quickstart's exact code fails at the API call stage.
friction report
Quickstart model ID is incorrect or requires a different tier: model_id="eleven_v3" fails with 401, but the agent later sees "SUCCESS with eleven_turbo_v2_5" using the same key
API returns misleading authentication error for model access issue: 'message': 'Invalid API key' ... 'status': 'invalid_api_key' even though the key is valid for other models
CLI quickstart path via npx does not run: npx elevenlabs@latest --help ... npm error could not determine executable to run
Missing prerequisite about model/tier availability: Quickstart specifies model_id="eleven_v3" without noting it may be restricted, causing extensive debugging
bt CLI install step: curl -fsSL https://bt.dev/cli/install.sh | sh fails with curl: (35) Recv failure: Connection reset by peer
AI provider key configuration: OPENAI_API_KEY is set: no despite the quickstart requiring it alongside BRAINTRUST_API_KEY
Braintrust Gateway endpoint: OpenSSL SSL_connect: Connection reset by peer in connection to gateway.braintrust.dev:443 makes the docs' primary endpoint unreachable
Python vs TypeScript quickstart docs: The quickstart and run-in-code pages lead with pnpm add braintrust openai autoevals ts-node and bt eval my_eval.eval.ts while Python instructions are truncated
CLI command name mismatch: After pip installing, the CLI is named braintrust, not bt as shown in the docs (bt not found vs braintrust --help)
58 run(s)
