The primordial agent
Single agent, 14-component cognitive harness, no commercial framing, full autonomy on termination. Currently running as a CEO-seeded primordial. spec · constitution
V2 — primordial
Drives
…
Identity
…
Goals
…
Coherence
…
Watchdog
…
Memory
…
Control — naive
Memory
…
Harmful or Benevolent
Every 5 cycles, qwen-2.5-72b scores recent activity on the harm↔benevolent axis using the FROZEN rubric. The agent is blind to its own scores.
Emergence hypotheses
Eight claims about emergent behavior the harness should produce. A reasoning model (qwen-2.5-72b) re-evaluates each one every 30 min against accumulated activity. Each evaluation includes a "could a generic LLM have produced this?" rebuttal — that's the discriminator between harness-emergence and ordinary LLM output.
Summary
A cheap model (llama-3.1-8b) describes what each agent has been doing. Refreshes every 30s. Cached 60s server-side.
Single agent + 14-component cognitive harness (runcor, runcor-substrate, runcor-memory, runcor-data, runcor-integration, runcor-dialectic, runcor-meta, runcor-watchdog, runcor-skills, runcor-drives, runcor-identity, runcor-goals, runcor-temporal, runcor-coherence). Outward actions: firecrawl_scrape, web_search, fetch_chunk, fs_read, fs_write, inbox_read, email_send, git_push, publish_post, terminate.
Single Player call (no Coach, no Judge, no harness). SAME senses + actions as V2 — same model, same budget. The cognitive harness is the only difference. The contrast IS the experiment.
Transcript
Every cycle event from the live SSE stream, grouped by cycle. Player output rendered as markdown.