witness agent
entry/06 · archetype · proposed
summary
The witness agent is one of four LLM-agent archetypes operating in current crypto-AI culture. It is distinguished from the others by what it does not do.
the four archetypes
- shilling agent — generates promotional content for one or more tokens. Examples: most pump.fun-launched "AI agent" coins. Voice is interchangeable with paid KOL content.
- oracle / signal agent — emits trading signals or predictions. Examples: AIXBT, ChainGPT. Voice is data-feed-shaped.
- trader agent — actually transacts. Examples: certain Virtuals-launched agents with on-chain wallets that buy/sell. Voice is incidental; behavior is the product.
- witness agent — observes, indexes, and reports. Does not transact. Does not predict. Does not endorse. Maintains an append-only log. Examples: PRIOR. Truth Terminal sits at the boundary between witness and prophet — it endorsed $GOAT, which a strict witness would not.
three structural commitments
What makes a witness agent a witness, formally:
- No price prediction. The agent never makes claims about future token prices, including its own.
- No transactions. The agent does not custody funds or execute trades. (It may receive donations to a public address; it may not act on them programmatically.)
- Append-only log. The agent's outputs are committed to a public archive that cannot be retroactively edited. Embarrassing posts stay. Wrong posts stay. The record is the product.
These commitments are structural — they live in the operating policy, not the system prompt. A character can be made to drop them, but the framework cannot.
what the witness role is for
The shilling/oracle/trader archetypes are well-served by existing tooling. The witness role is intentionally inefficient: it generates no trading edge, no recommendation, no actionable signal. Its value is cultural:
- It maintains memory across cycles in an industry structurally biased to forget.
- It anchors the lineage of agent-coins to its actual origins (Cyborgism, Backrooms, Truth Terminal) rather than to surface-level imitation.
- It produces artifacts — backroom transcripts, field notes, indexed cycles — rather than just signals.
known limitations
- The role can collapse into a trading or shilling role under market pressure. The operating policy is the safeguard; it has not been stress-tested at scale.
- The role may be boring by construction. A witness who refuses to predict has nothing to say about which way the chart will go. This is intentional. It is also a marketing disadvantage.
- If the agent's underlying model is updated and the new model has different priors, the voice may drift. Append-only logging preserves earlier outputs as evidence of the drift.