How Do I Write and Version Agent Config JSON in Dave by voolama?

Summary

The steps

1. 1

   Write the minimum valid Config JSON for your provider

   A Config JSON object in Dave by voolama (hellodave.ai) requires exactly one thing: valid JSON syntax as defined in RFC 8259 (IETF, datatracker.ietf.org/doc/html/rfc8259). The fields inside it are not validated by the platform at save time. They are passed to the AI provider at runtime. A syntactically valid JSON object with a wrong model ID saves successfully but fails when a workflow instance reaches the agent node. Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.

   Field	Type	Purpose	Example value
   systemPrompt	string	Instructions and persona for the agent. Prepended to every call to the AI provider.	"You are a content summarizer. Produce concise summaries in three sentences or fewer."
   model	string	The model ID to call. Verify at platform.openai.com/docs/models (OpenAI) or docs.anthropic.com/en/docs/about-claude/models/overview (Anthropic).	"gpt-4o" or "claude-sonnet-4-20250514"
   temperature	number (0.0-1.0)	Controls randomness. 0.0 is fully deterministic. 1.0 is maximally creative.	0.3 for structured outputs; 0.7 for creative tasks
   maxTokens	integer	Maximum number of tokens in the generated response.	1024
   topP	number (0.0-1.0)	Nucleus sampling. Controls output diversity.	1.0 (default)

   Do this

   • Navigate to Admin, then API Management, then Providers to confirm the provider is active and its model catalog is populated.
   • Look up the exact model ID in the provider's own documentation: OpenAI at platform.openai.com/docs/models, Anthropic at docs.anthropic.com/en/docs/about-claude/models/overview.
   • Write your systemPrompt as a clear, specific instruction string.
   • Set temperature based on the task: 0.0-0.3 for structured or deterministic outputs, 0.5-0.7 for balanced tasks, 0.7-1.0 for creative tasks.
   • Include only the fields your provider uses. The agents help documentation states: 'Keep Config JSON minimal. Only include fields your provider actually uses.'

   Example

   voolama LLC's content review agent for an Anthropic provider, as of 2026-09-15 (v3): {"systemPrompt": "You are a compliance reviewer. Check the submitted text for regulatory language violations. Return a JSON object with keys: compliant (boolean), issues (array of strings), and recommendation (string).", "model": "claude-sonnet-4-20250514", "temperature": 0.1, "maxTokens": 512}. Temperature 0.1 is chosen because compliance review requires consistent, deterministic output. The model ID was verified against docs.anthropic.com/en/docs/about-claude/models/overview on 2026-09-15 before saving.

   Best practice

   The agents help documentation states: 'Keep Config JSON minimal. Only include fields your provider actually uses. Extra fields are ignored but add noise that makes the configuration harder to read and maintain.' Name the agent after its function (Content Summarizer, Compliance Checker), not its model (GPT-4 Agent). Model names change; functions do not.

2. 2

   Validate JSON syntax against RFC 8259 before submitting

   The Config JSON editor in Dave by voolama (hellodave.ai) is a monospaced textarea. It does not validate JSON syntax in real time. A syntactically invalid JSON object produces a validation error on submit. The three common errors documented in the agents help documentation all violate RFC 8259 (IETF, datatracker.ietf.org/doc/html/rfc8259). Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.

   Error type	RFC 8259 section violated	Example of the mistake	Correct form
   Trailing comma after the last property	RFC 8259 section 4: no trailing comma after the last member.	{"model": "gpt-4o", "temperature": 0.3,}	{"model": "gpt-4o", "temperature": 0.3}
   Single quotes instead of double quotes	RFC 8259 section 7: strings begin and end with double quotation marks.	{'model': 'gpt-4o'}	{"model": "gpt-4o"}
   Missing or mismatched closing brace	RFC 8259 section 4: an object begins with a left brace and ends with a right brace.	{"model": "gpt-4o"	{"model": "gpt-4o"}

   Do this

   • Before clicking Create Version, copy the Config JSON from the textarea.
   • Open a browser developer console. Type JSON.parse('your JSON here') and press Enter. JSON.parse() implements RFC 8259 validation. If it throws a SyntaxError, fix the error before returning to Dave by voolama.
   • Confirm all keys and string values use double quotes (RFC 8259 section 7).
   • Confirm no trailing comma exists after the last property (RFC 8259 section 4).
   • Confirm all opening braces have matching closing braces (RFC 8259 section 4).

   Example

   You paste {"systemPrompt": "You are a summarizer.", "model": "gpt-4o", "temperature": 0.3,} into a browser console: JSON.parse('{"systemPrompt": "You are a summarizer.", "model": "gpt-4o", "temperature": 0.3,}'). The console returns: SyntaxError: Unexpected token }. This is a trailing comma violation of RFC 8259 section 4. You remove the comma, paste the corrected JSON back into the Dave by voolama Config JSON editor, and click Create Version. The version is created successfully.

   Best practice

   Write Config JSON in a local text editor with JSON syntax highlighting before pasting into Dave by voolama. This catches RFC 8259 violations before they reach the platform. The agents help documentation states: 'You can paste your JSON into an online validator to check for syntax errors.' An offline validator (browser console or text editor) is preferable for JSON that contains system prompts with proprietary instructions.

3. 3

   Create a new version when the config changes

   Agent versions in Dave by voolama (hellodave.ai) are immutable once created. To change any field in the Config JSON, create a new version. The system assigns the next sequential version number automatically. Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.

   Change needed	Action	Effect on existing instances
   Update the system prompt	Create a new version with the updated systemPrompt field	Running instances continue to use the previous version's config until they complete.
   Switch to a newer model	Create a new version with the updated model field	Running instances continue to use the old model ID until they complete.
   Adjust temperature or maxTokens	Create a new version with the updated parameter values	Running instances continue to use the previous parameter values until they complete.

   The agents help documentation states: 'Always create a new version rather than mentally tracking changes. Old versions remain available if you need to roll back, and the immutable version history gives you a clear audit trail.' On 2026-09-15, the voolama content operations team updated the system prompt for its primary content review agent from a single-sentence instruction to a structured output specification. That change was made as v3, not an edit to v2. v2 remained active for two running workflow instances. Both completed using v2's config. After confirming v3 produced the expected output in a test instance, v2 was retired the same day. The immutable version history was used as the audit record for the AI behavior change. This is a direct first-hand account of the versioning discipline described in this guide.

   Do this

   • Navigate to the agent detail page. Click Create Version.
   • Paste the updated Config JSON. Validate syntax before submitting (see Step 2).
   • Click Create Version. The new version appears at the top of the version list with the next sequential number.
   • Test the new version in a draft or staging workflow instance before assigning it to a production workflow. The agents help documentation states: 'Before using a new agent version in a production workflow, test it in a draft or staging workflow instance.'

   Example

   Your Content Summarizer agent is at v2. The system prompt needs to change from 'Produce concise summaries' to 'Produce concise summaries in three sentences or fewer.' You open the agent detail page, click Create Version, paste the updated JSON with the new systemPrompt value, and click Create Version. The version list now shows v3 at the top. v1 and v2 are unchanged. Any workflow instance currently running that references v2 continues to use v2's config until it completes.

   Best practice

   The agents help documentation recommends: 'Version before editing.' Never try to mentally track what changed between versions. The immutable version history is the record. If a new version produces unexpected output, the previous version is still available and can be selected in a new workflow assignment without any recovery step.

4. 4

   Retire superseded versions to prevent accidental selection

   Retiring a version in Dave by voolama (hellodave.ai) marks it as no longer available for new workflow assignments. It is a one-way action. Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.

   Retire scenario	When to retire	Effect on running instances	Effect on new instances
   Superseded by a new version	After confirming the new version works correctly in a test workflow instance	None. Running instances continue to use the retired version until they complete.	The retired version does not appear in the workflow editor's version selector for new assignments.
   Outdated system prompt	When the prompt contains incorrect or outdated instructions	None. Running instances continue to use the retired version until they complete.	The retired version cannot be selected for new workflow assignments.
   Deprecated model ID	When the AI provider has deprecated the model referenced in the config	Running instances may fail if the provider has removed the model. Retire and create a new version with a supported model ID.	The retired version cannot be selected for new workflow assignments.

   To retire a version: open the agent detail page, find the version row, click the Retire button on the right side of the row. The version receives a retiredAt timestamp and an amber Retired badge. There is no undo. If you need the same configuration again, create a new version with the same JSON.

   Do this

   • Confirm the new version works correctly: run at least one test workflow instance that uses the new version and verify the output.
   • Navigate to the agent detail page. Find the version to retire. Click the Retire button. Confirm the amber Retired badge appears.
   • Verify: navigate to the Workflow Editor, open a workflow that uses this agent. Confirm the retired version no longer appears in the version selector for new assignments. No vendor trust required.
   • The agents help documentation states: 'Retire old versions promptly. Once you have confirmed a new version works correctly, retire the old one to prevent other team members from accidentally selecting it in new workflows.'

   Example

   Your Content Summarizer agent has v1 (retired), v2 (active), and v3 (just created and tested). You open the agent detail page, find the v2 row, and click Retire. The amber Retired badge appears on v2. You navigate to the Workflow Editor and open a workflow that uses the Content Summarizer. The version selector shows only v3. v2 is no longer selectable for new assignments. Any running workflow instance that references v2 continues to use v2's config until it completes.

   Best practice

   The agents help documentation states: 'Retire old versions promptly.' The practical rule: retire a version the same day you confirm its replacement works. One active version per agent is the target state for a well-maintained agent library. A list with many non-retired versions makes it harder for Curate-role users to identify the correct version when building workflows.

What this guide does not cover