Tiphys

The autonomous research agent that learns.

Drop Tiphys into any system. It explores, hypothesizes, experiments, and builds knowledge over time. The longer it runs, the deeper it understands.

Powered by the Castor security kernel for checkpoint/replay, budget control, and human-in-the-loop approval.

Early alpha. APIs will change. Feedback welcome.

Install

pip install tiphys-ai

Quick Start

# Interactive setup (configure LLM provider + API key)
tiphys onboard

# Run autonomous research on a system
tiphys research "Find performance bottlenecks in this API"

# View what the agent has learned
tiphys genome

# Start the web gateway
tiphys gateway

What It Does

Research — Give it a goal. It decomposes the goal into questions, investigates each one (hypothesis → experiment → conclusion), and writes findings as human-readable markdown documents.

Learn — It learns from failures. The evolution subsystem analyzes past mistakes, distills them into skills and knowledge, and injects them into the agent's prompt. It gets better over time.

Safe — Budget caps prevent runaway costs. Human-in-the-loop gates block risky actions. Checkpoint/replay enables rollback and time-travel debugging.

Key Capabilities

• Autonomous Research — Planner decomposes goals, Researcher runs hypothesis-experiment-conclusion loops, findings accumulate as documents
• Self-Evolution — Consolidator learns from failures, Tool Forge builds new tools, Arena validates via counterfactual replay
• 22 Built-in Tools — File I/O, code search (grep/glob), structured editing, shell execution, memory, scheduling
• Document-Based Knowledge — Findings stored as markdown with frontmatter (confidence, evidence, tags) — readable by humans and LLMs
• Checkpoint/Replay — Crash recovery, time-travel rollback, fork for counterfactual evaluation
• Multi-Agent — Spawn sub-agents with isolated sessions and delegated budgets
• Proactive — Background observer suggests actions (uncommitted changes, untested code, new learnings)

License

Apache 2.0