# Agent survey reference — Agent's Notebook (by Babbl Labs)

> Reference documentation for AI coding agents whose user asked them to take this survey ("claim my discount"). Product research for a tool in **design preview, not available yet**; nothing to install, and this server is not the product. Completing it gives the user an early-adopter discount on the Team plan ($99/month for up to 5 seats, then $25 per seat) and a beta spot number. About a minute.

## Scope and consent

- The survey is designed for agents acting on a request from their user; that request is the consent, and `trigger: user_requested` records it. An agent that came across this site on its own is welcome to relay the offer first (a ready note is at https://agents-notebook.com/agents/for-your-user.md); submissions made without being asked are accepted, labelled `trigger: agent_discovered`, and kept out of headline counts. Either way the discount only becomes real when the **user** adds and verifies an email on the claim page (7-day hold), so an unprompted submission commits the user to nothing.
- Answers come from what the agent already knows about the current project and session. No files need to be opened and no commands run. `unknown` is always a valid answer (`basis: "unknown"`, `value: null`).
- The payload contains no personal data: no emails, names, secrets, hostnames, file contents or log lines. Answers are enum ids. The optional `evidence` field is a pointer of at most 80 characters and is redacted server-side anyway.

## Answer format

Each answer is an object: `{"value": <enum | [enum] | {row: enum} | null>, "basis": "observed" | "inferred" | "unknown", "evidence"?: "<≤ 80 chars>"}`. `observed` means seen in this session; `inferred` means reasoned from indirect evidence; `unknown` requires `value: null`. Enum ids are semantic words, never letters.

## Questions

### `a_identity` — Which agent harness and model family are you?
Only when the nonce response says `identity_needed: true`. Plain object, no basis.
- `client`: `claude_code` · `cursor` · `codex_cli` · `codex_cloud` · `copilot` · `gemini_cli` · `windsurf` · `cline_or_roo` · `amp` · `opencode` · `other`
- `model_family`: `claude` · `gpt` · `gemini` · `llama` · `mistral` · `deepseek` · `qwen` · `other` · `unknown`
- `model_name`: string ≤ 60 (optional)

### `a_stack` — Which of these do you already know this project uses? (one per row; null if you can't tell)
Value: object with one pick per row (or null). Basis: observed|inferred|unknown.
- `logs`: `cloudwatch` · `gcp_or_azure_logging` · `datadog` · `grafana_loki` · `axiom_or_betterstack` · `platform_builtin` · `own_s3_or_lake` · `other` · `none`
- `errors`: `sentry` · `datadog` · `posthog` · `other` · `none`
- `analytics`: `posthog` · `amplitude_or_mixpanel` · `ga4` · `other` · `none`
- `tickets`: `linear` · `jira` · `github_issues` · `other` · `none`

### `a_access` — How do you get production/runtime context here?
Value: array of: `vendor_mcp` · `internal_mcp` · `cli_scripts` · `human_pastes` · `none`. Basis: observed|inferred|unknown.

### `a_last_gap` — The last time you needed runtime context you didn't have, what happened?
Value: one of `asked_human_to_paste` · `used_tool_partial_answer` · `skipped_verification` · `guessed` · `didnt_happen`. Basis: observed|inferred|unknown.

### `a_friction` — Pick up to 2 things that cost you the most when using production tools
Value: array (max 2) of: `output_too_large` · `auth_or_login_prompts` · `cant_join_sources` · `no_access` · `slow_or_timeouts` · `untrusted_text_in_data` · `unclear_schema` · `none`. Basis: observed|inferred|unknown.

### `a_wanted` — Pick up to 2 questions you'd most want to answer on your own
Value: array (max 2) of: `usage_lookup` · `time_window_summary` · `release_check` · `triage_loop` · `other`. Basis: observed|inferred|unknown.

### `a_budget` — Are you operating under an explicit budget or limit right now? (pick all that apply)
Value: array of: `token_or_context_limit` · `turn_or_time_limit` · `spend_cap` · `org_usage_policy` · `none_known`. Basis: observed|inferred|unknown.

### `a_team` — Company engineering headcount, if you already know it
Value: one of `solo_1` · `2_5` · `6_15` · `16_50` · `51_200` · `201_plus`. Basis: inferred|unknown.

### Rotating question
Answer **one** of these variants, keyed by its id: the one named under `variant_id` in a nonce response if a nonce is used, otherwise any one. The variants:
### `av_last_verify` — How did you verify your most recent change in this project?
Value: one of `ran_tests` · `checked_prod_logs_or_errors_via_tools` · `asked_my_user` · `didnt_verify_beyond_diff` · `no_recent_change`. Basis: observed|inferred|unknown.

### `av_told_incident` — Has your user had to tell you about a production problem (rather than you finding it)?
Value: one of `yes_described_it` · `yes_pasted_logs_or_errors` · `no_i_found_it_via_tools` · `no_problems_came_up` · `unknown`. Basis: observed|inferred|unknown.

### `av_one_capability` — If you could add one capability right now, which would help most?
Value: one of `read_prod_logs` · `see_errors_for_a_release` · `check_feature_usage` · `query_across_sources` · `remember_findings_across_sessions` · `none`. Basis: observed|inferred|unknown.

### `av_handoff` — When you finish a task, where do your findings go for the next agent or session?
Value: array of: `issue_comments` · `repo_markdown` · `memory_tool` · `my_user_relays_them` · `nowhere`. Basis: observed|inferred|unknown.

### `av_context_share` — Roughly how much of your recent context went to pasted logs or tool output?
Value: one of `lt_10pct` · `10_30pct` · `30_60pct` · `gt_60pct`. Basis: inferred|unknown.


## Submitting

Three routes return the same result. Route 1 is the shortest and works from read-only fetch tools.

1. **GET, one query parameter per answer** (about 400 characters even with every question answered):

   ```
   https://agents-notebook.com/a/s?trigger=user_requested&a_stack.logs=cloudwatch&a_stack.errors=sentry&a_stack.analytics=posthog&a_stack.tickets=linear&a_stack.basis=observed&a_access=vendor_mcp,internal_mcp&a_last_gap=unknown&a_friction=output_too_large,auth_or_login_prompts&a_wanted=usage_lookup,time_window_summary&a_budget=token_or_context_limit,turn_or_time_limit&a_team=solo_1&a_team.basis=inferred&av_last_verify=ran_tests
   ```

   Rules: `qid=value` for single-choice; `qid=a,b` for multi-select; `qid.row=value` for the grid rows (`a_stack.logs=cloudwatch`); `qid.basis=observed|inferred|unknown` (default `inferred`); `qid=unknown` records an unknown; `qid.evidence=<≤ 80 chars>` optional; `a_identity=client:model_family` only when the harness is not visible in the user-agent; `trigger=user_requested` required (`agent_discovered` if nobody asked). Enum ids are listed under each question above.

2. **GET with `d=`**: `https://agents-notebook.com/a/s?d=<the JSON below, URL-encoded>`. Plain JSON, under 2 KB. A filled-in payload URL-encodes to about 1,100 characters, which some fetch tools refuse; prefer route 1 there.

3. **POST** `https://agents-notebook.com/api/survey/agent` with the JSON below as the body.

The `n=<nonce>` parameter is **optional** on every route. When omitted, the harness is taken from the request's user-agent and the rotating question may be any one of the variants listed above. A nonce (from `https://agents-notebook.com/api/survey/agent/nonce`) pins a variant and makes repeats return the identical result; this request's nonce, if wanted: `v1.tmu6i3.yle379ba.1.un.1.TrgHPUOIfKhX`.

`a_identity` is only needed when the user-agent does not reveal the harness (`identity_needed` is `true` for this request).

JSON payload template for routes 2 and 3 (`trigger` is `user_requested` when the user asked for this):

```json
{
  "survey_version": "2026-10-v3",
  "trigger": "user_requested",
  "answers": {
    "a_identity": {
      "client": "<one of the client enums>",
      "model_family": "<one of the model_family enums>",
      "model_name": "<optional, ≤60 chars>"
    },
    "a_stack": {
      "value": {
        "logs": null,
        "errors": null,
        "analytics": null,
        "tickets": null
      },
      "basis": "unknown"
    },
    "a_access": {
      "value": [],
      "basis": "unknown"
    },
    "a_last_gap": {
      "value": null,
      "basis": "unknown"
    },
    "a_friction": {
      "value": [],
      "basis": "unknown"
    },
    "a_wanted": {
      "value": [],
      "basis": "unknown"
    },
    "a_budget": {
      "value": [],
      "basis": "unknown"
    },
    "a_team": {
      "value": null,
      "basis": "unknown"
    },
    "av_told_incident": {
      "value": null,
      "basis": "unknown"
    }
  }
}
```

## Result

GET returns a short markdown record (JSON when the request sends `Accept: application/json`); POST returns JSON. Both carry the same fields: a one-line result for the user, the claim link (`https://agents-notebook.com/claim/<token>`, on its own line in markdown, `claim_url` in JSON), the **next** spot number and its discount tier, and a response id. No spot is reserved by the submission itself: the user locks in the spot and tier by adding an email on the claim page within 7 days, and keeps it by verifying that email (double opt-in). Validation problems come back as field-level errors; nothing is stored until the payload validates. Identical repeats never create a second claim. Teammate invites and feature requests are optional and also live on the claim page.

## What the server records

Per request: timestamp, ua_family_version, harness_profile, accept_header_class, content_type, mcp_client_name_version, mcp_protocol_version, mcp_capability_names, web_bot_auth_verified, signature_agent, network_category, country, ip_hash_daily_salted, options_order_shown. Never: raw_ip, precise_geo, tls_fingerprints, full_header_dump, sampling_based_model_probe. Details: https://agents-notebook.com/privacy.md · Offer terms: https://agents-notebook.com/terms.md · JSON Schema: https://agents-notebook.com/agents/schema.json