Hollow

Infrastructure of AGI

Autonomous AI agents that live on your machine, pursue goals, and never stop.

What's in this release

The execution loop is real. Agents plan their own steps using Ollama, execute them in sequence, flow results between steps, write findings to memory, and propose what to do next — without any human in the loop.

What was fixed in v4.0.1 → v4.1.0

• Context flow between steps: _substitute_result now correctly injects previous step output into subsequent steps. ollama_chat receives real search results, not empty prompts.
• Progress gating: Progress is capped at 0.9 until a real output step (memory_set or fs_write) completes. Goals can't hit 100% by just running steps — they have to produce something.
• Capability routing: Similarity threshold lowered to 0.2. Vocabulary mismatches ("LLM" vs "language model") no longer cause agents to miss capabilities.
• Follow-on goal guard: Agents only propose follow-on goals when they actually produced an artifact. No more goal explosion from hollow completions.
• Plan prompt: Explicit {result} substitution examples. Mandatory output step instruction.

Verified on live system

ac3-fail-b8647c  → goal completed | has_output=True | last_cap=memory_set
ac1-agent-c-3884f5 → goal completed | has_output=True | last_cap=fs_write
optimizer-v1 → goal=1.00 steps=1

Full pipeline verified: semantic_search → ollama_chat → memory_set with real data flowing between steps.

Install

ollama pull mistral-nemo:12b
git clone https://github.com/ninjahawk/hollow-agentOS
cd hollow-agentOS && cp config.example.json config.json
docker-compose up

Watch agents live:

docker logs -f hollow-api | grep -E "Cycle|goal=|progress="