07 AI market analyst · Instructor
Run the agent module so the room referees the answer instead of applauding it.
macOS terminal: Run workshop commands in Terminal using zsh or bash.
Timing
20 minutes. Spend 5 on agent creation and instructions, 10 on the three turns, 5 on the referee query and the wrong-answer sweep. Do not let the room free-chat past minute 15.
Talk track
- ClickHouse Agents runs on the learner's existing Cloud identity. No new key, no local process, no model subscription.
- The agent instructions are the lesson. Each rule maps to a real failure: aggregate states
need
Merge,marketsneedsFINAL, the newest minute is partial, and last-trade ticks are not probabilities. - Turn one is detection, turn two is investigation, turn three is self-attack. Only the third turn tells you whether the verdict was reasoned or narrated.
- Fixture mode is fine here. Say out loud that a fixture move is synthetic, and expect the agent not to know that; it will describe a pattern as market behaviour.
- The referee query is authoritative because it is deterministic. When the two disagree, the SQL is right until someone shows otherwise.
- Guardrail: public data, indicative midpoints, no trading advice. If an agent offers a recommendation, that is a finding for Step 6, not a joke.
Common failures
- ClickHouse tool shows disconnected: reopen the tool, select
polymarket-workshop, and confirm before any prompt. A disconnected agent invents plausible column names. - Agents unavailable in the organization: pair with a learner who has it, or run the module through the ClickHouse MCP from a coding agent using the same prompts.
- Empty detection result: the aggregate needs six minutes of data. Check collector health
and
max(minute)before debugging the prompt. - Agent selects raw
AggregateFunctioncolumns: expected. Have the learner paste the error back and record it in Step 6 rather than fixing the prompt for them. - Yes and No reported as two movers: ask which condition ID both tokens belong to.
- Learner accepts the answer without the referee query: the module is not complete without a written-down mistake.
Verification
Ask each learner for three things: the detection SQL the agent showed, the top row of the referee query, and one sentence on what the agent got wrong or overstated. All three, or the module is not done.
Reset
No state changes. Delete the agent and recreate it if the instructions were edited into something unrecoverable:
Discard this conversation and start again from the agent instructions.Nothing in ClickHouse Cloud needs repair; the agent only ran SELECT queries.