Poggio AI

This page covers the main workflows for using pAI MSc effectively.

Running Research

Using presets

# Quick
msc run "What are the key differences between transformer and state-space models?" --preset quick

# Standard
msc run "Survey the landscape of mechanistic interpretability methods"

# Thorough
msc run "Analyze the theoretical foundations of in-context learning" --preset thorough

# Maximum
msc run "Comprehensive analysis of attention mechanisms" --preset maximum

Using a task file

msc run --task-file my_task.txt

Direct pipeline invocation

python launch_multiagent.py \
  --task "Investigate whether batch normalization reduces spectral norm growth" \
  --enable-counsel \
  --enable-math-agents \
  --enforce-paper-artifacts \
  --min-review-score 8

Output directory

All output is written to results/consortium_<timestamp>/.

File	Description
`final_paper.tex` / `final_paper.md`	Generated manuscript
`final_paper.pdf`	Compiled PDF
`run_summary.json`	Cost, tokens, stages completed
`budget_state.json`	Cumulative spend by model
`paper_workspace/`	Research workspace artifacts
`math_workspace/`	Proof/verification artifacts (if enabled)
`STATUS.txt`	`COMPLETE`, `INCOMPLETE`, or `ERROR`

Checking Status and Logs

msc status
msc logs -f
msc runs

Resuming Interrupted Runs

msc resume

Or:

python launch_multiagent.py \
  --resume results/consortium_20260307_120000/ \
  --start-from-stage writeup

See Configuration for stage aliases and flags.

Counsel Mode

msc run "Your research question" --preset thorough

python launch_multiagent.py \
  --task "Your research question" \
  --enable-counsel \
  --counsel-max-debate-rounds 3

Counsel mode requires API keys for multiple providers and costs about 4x a single-model run.

Math Agents

python launch_multiagent.py \
  --task "Prove convergence bounds for gradient descent on L-smooth functions" \
  --enable-math-agents

Tree Search

python launch_multiagent.py \
  --task "Your research question" \
  --enable-tree-search \
  --tree-max-breadth 3 \
  --tree-max-depth 4 \
  --tree-max-parallel 6 \
  --tree-pruning-threshold 0.2

Campaigns

msc campaign init --name "my_project" --task "Investigate normalization layers in transformer training dynamics"
msc campaign start my_project_campaign.yaml
msc campaign status my_project_campaign.yaml
msc campaign list

Budget Management

msc budget
msc config set budget_usd 50

Alerts trigger at 85%, 95%, and 100%.

Notifications

msc notify setup
msc notify test --channel telegram

OpenClaw Integration

msc openclaw setup
msc openclaw start
msc openclaw status

Adversarial Verification

python launch_multiagent.py \
  --task "Your question" \
  --adversarial-verification

Usage