Autodialectics
https://github.com/hmbown/autodialectics- Category
- AI Agents
- Rank
- No. 1792Tools index
Previous survey · No. 1731 ·
- Pricing
- Open Source
- Type
- AGENT
- Builder
- Hmbown
- GitHub
- 4 stars
- Date
About
Agentic harness for keeping research-oriented LLM runs on-task with contracts, evidence, dialectics, verification, and anti-slop gates.
What it does
Autodialectics turns a task description into a fixed specification, gathers relevant source material, challenges its initial plan, executes through a domain adapter, then checks the result before accepting it. It stores intermediate artifacts and supports replay, benchmarking, and policy comparison.
Why it's ranked here
The workflow addresses recognizable agent failures with explicit stages and inspectable artifacts. Code tasks can copy a workspace into a sandbox, run specified checks, and retry within a fixed budget. The verdict stays mixed because important verification and benchmark paths still rely on brittle text overlap and phrase matching.
What's good
Contracts preserve objectives and forbidden shortcuts. Thesis, antithesis, and synthesis require objections before execution. Verification runs separately from execution, while artifacts and manifests preserve what happened. Code workflows support explicit workspace roots, user-supplied verification commands, bounded repairs, and failure feedback. Optional DSPy paths fall back to heuristics when unavailable.
Tradeoffs
Several safeguards are heuristic rather than semantic. Verification uses keyword overlap in important fallback paths, while benchmark checks can depend on literal phrases. Repository notes identify false positives around quoted warnings and completion claims. Stricter sandbox isolation, stronger live endpoint coverage, and hardened DSPy benchmarking remain stated priorities.
How to use it well
Use it for research, writing, analysis, experiments, or repository changes where traceability and explicit acceptance checks justify a multi-stage run. Supply focused assets, an explicit domain, and concrete verification commands. Treat its scores as signals requiring inspection. It does not provide the planned richer model registry, routing telemetry, hooks, or strict sidecar isolation.
Technical notes+
pyproject.toml defines a Python 3.11+ Hatchling package with Typer CLI, FastAPI, MCP, httpx, Pydantic, SQLite-facing runtime dependencies, optional DSPy, and pytest extras. autodialectics/api/app.py exposes health, task compilation, run, benchmark, evolution, promotion, and rollback routes; it stores a module-level runtime instance and initializes it on the first dependency request. autodialectics/settings.py resolves explicit, environment, project, then user configuration and defaults to a local OpenAI-compatible endpoint. autodialectics/utils/text.py implements lexical chunking, token overlap, and repetition measures. docs/linear_handoff_2026-03-26.md explicitly records remaining phrase-matching, overlap-heavy verification, domain-inference, provider-coverage, and false-positive risks.
Observed
- License
- MIT
- Primary language
- Python
- Runtime requirement
- Python 3.11 or newer
- Packaging
- Hatchling wheel package, installed and run through uv; optional dev and DSPy extras
- Interfaces
- Typer CLI, FastAPI REST API, and MCP server
- Model endpoint
- OpenAI-compatible chat completions endpoint
- Persistence
- JSON and Markdown artifacts tracked with SQLite
- Repository structure
- Includes tests, benchmarks, documentation, examples, and local integrations for Codex, Claude Code, and OpenCode
Read from README.md, pyproject.toml, docs/architecture.md, docs/reference_flow.md, docs/ai-plugin-integrations.md, docs/cliproxy_sdk_extension.md, docs/linear_handoff_2026-03-26.md, autodialectics/__init__.py, autodialectics/settings.py, scripts/validate_ai_integrations.py, autodialectics/api/app.py, autodialectics/cli/main.py, autodialectics/utils/dspy.py, autodialectics/utils/text.py.
What it can do
Enforce research contracts on LLM runs
Research contract parameters and LLM query → Contract-compliant LLM response
Validate evidence in LLM outputs
LLM-generated content with claimed evidence → Evidence verification report
Execute dialectical reasoning processes
Research question or argument → Structured dialectical analysis
Filter low-quality content through anti-slop gates
LLM-generated text → Quality-filtered content
Keep LLM research runs focused on specified tasks
Task definition and LLM interaction parameters → On-task research results
Verify accuracy of research-oriented LLM outputs
LLM research findings → Verification status and accuracy assessment
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.