Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Using LLMs for iteratively exploring the solution search space at scale.
| Date | Stars |
|---|---|
| 2026-07-31 | 738 |
| 2026-08-05 | 740 |
| 2026-08-06 | 740 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Iterative Studio
The system integrates with major AI providers (Google AI, OpenAI, Anthropic) and employs multi-agent-based architectures. The system is capable of running with local models in fully offline mode.
## Operational Modes
The system operates in three distinct modes, each optimized for specific use cases.
### 1. Deepthink Mode
**Purpose**: High-depth problem solving through independent strategic branches, targeted hypothesis testing, parallel execution, critique, correction, and explicit final selection.
Deepthink supports two execution families:
1. **Single-pass strategic search**: Generates up to ten main strategies, optionally expands each into sub-strategies, tests hypotheses, executes all branches, and optionally performs critique synthesis, full-solution context correction, or both.
2. **Evolving Depth First Search**: Runs up to five direct strategy branches through repeated correction and critique. Each branch has a structured breadth-first solution pool, recursive memory bank, selective hypothesis packet, and periodic Post Quality Filter evaluation.
In Evolving DFS, the original execution is iteration 1. Subsequent iterations correct and critique the active solution, refresh strategy-specific hypotheses every two global iterations, and run memory/PQF maintenance after each branch accumulates five new history entries. PQF can keep a branch or replace its strategy in the same stable slot with a clean, versioned branch.
The Structured Solution Pool is the BFS companion to the depth-first correction loop. It creates five substantively executed alternatives per strategy and iteration, while correctors receive deep local history and only limited cross-strategy context. Replaced branches remain archived for inspection but are excluded from active prompts and final judging.
Hypothesis routing supports Blind Trust, Strategy-Aware, and Selective modes. Evolving DFS always uses Selective mode and injects each strategy's tested packet into its execution, correction, and solution-pool agents.
The Final Judge sees only active candidate solution texts. It does not receive critiques, memory banks, solution pools, PQF decisions, or replaced branches.
**Sandbox Environment and Artifact Submission**:
Deepthink integrates a secure sandbox virtual environment for execution and verification.
- **Repository Visibility**: Every Deepthink role receives `sandbox_exec` and `final_output` when the Sandbox Terminal Environment is enabled. Active branches use `Strategy-N/{Critique,SolutionPool}`: execution and correction write direct branch files, critique owns `Critique`, and the pool owns `SolutionPool`. PQF replacements archive the complete old branch under `Pruned_Strategies/Strategy-N_First_PQF` (then ordinal successors) before recreating fresh active slot directories. Hypothesis tests are organized by `Hypothesis-vN`, while only current selectively routed tests are mounted to branch workers.
- **Submit Final Artifact**: Sandbox-enabled agents use `sandbox_exec` for iterative exploration and testing, then use `final_output` to submit their completed work. JSON-producing roles submit their existing role-specific JSON object directly through `final_output`; the environment validates that contract in the tool loop and returns a correction error without discarding the agent's research. Downstream agents and the central system receive only the submitted artifact, filtering out intermediate command transcripts and scratchpad data.
See [Deepthink architecture and context flow](Deepthink/DeepthinkDocs.md) for the complete agent contracts, repository schemas, mode behavior, iteration synchronization, and failure policy. The previous diagram remains archived at `Deepthink/OldSystemArchitecture.png`.
### 2. Adaptive Deepthink Mode
**Purpose**: An orchestrator-directed, pass-based Deepthink workflow for divergent strategic search without a separate final jExcerpt of 9,972 characters
Read on GitHub
96
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:ffc5602cb3098abc, llm:description: 'Using LLMs for iteratively exploring the solution search space at scale.' (repository description)
matched fp:ffc5602cb3098abc, llm:description: 'Using LLMs for iteratively exploring the solution search space at scale.' (repository description)