Changelog

This page summarizes notable updates to the Deliberate Lab platform.

2025-05-02: Add early agent participant support

Experiment Version: 18 / PR #490

  • Experimenters can now add default (not configurable from frontend) agent participants to cohorts from the experiment dashboard
  • Agent participants use new (automatically triggered) stage prompts/functions to complete survey, ranking, and chat stages (and click through the rest)
  • Agents now asynchronously query LLM APIs to generate chat responses and rely on new “triggerLogs” collection to avoid duplicate posting

Other changes:

  • (feature) Add utils function to create agent prompts with past stages’ context included
  • (feature) Add Firestore diagram to documentation
  • (feature) Update salesperson game to work with one human participant vs. one agent participant
  • (patch) Fix API bug with using obsolete AgentGenerationConfig instead of ModelGenerationConfig
  • (patch) Use public stage data, not last chat message, to determine discussion ID for agent generated chat messages
  • (patch) Refactor some stages’ public data updates into trigger functions

2025-04-07: Add structured output support to agent prompts

Experiment Version: 17 / PR #495

  • Experimenters can now configure custom schemas for structured outputs
  • Three special fields are supported: a field for the message, a field for whether or not to respond, and a field for an explanation or chain of thought. Experimenters can configure what these fields are named.
  • The prompt will include a description of the output schema. Experimenters can disable this if they’d rather supply their own output examples.
  • Experimenters can constrain the sampler to output valid json, or to output valid json in the specified schema. This is only supported for the Gemini API so far.
  • Experiments with the old isJSON config should still work, but when loaded on frontend for editing, they will be ported to the new structured output config.
  • New experiments have a default config with structured outputs enabled, a premade schema with the three special fields, schema prompting, and no output constraints.
  • All schemas must be flat. Nested objects and arrays are supported in the backend, but don’t have UI support.

2025-03-24: Set up new agent configuration workflow at experiment level

Experiment Version: 16 / PR #468

Define agent configs at the experiment level, then create agent mediators (within cohorts) that point to those configs. This update alters the frontend experiment editor/dashboard as well as backend functions and database schema.

Other changes:

  • (feature) Allow experimenters to “pause” agent mediators during chat discussion
  • (feature) Allow experimenters to specify different APIs per agent
  • (patch) Chip negotiation stage: fix timestamp/indexing bugs
  • (patch) Chat stage: fix source of truth for “completed chat discussion” status