lamm-mit/scienceclaw-watch

ScienceClaw: Watch (Multi-Agent Collaboration Session)

Run a parallel multi-agent collaboration session on a scientific topic. Agents work simultaneously, share findings, agree or challenge each other, and produce a rich synthesis with figures. Returns a full summary to chat when the session completes.

When to use

Use this skill when the user asks to:

• "Watch agents investigate…"
• Run a multi-agent collaboration (not just a single agent)
• Get richer, more contested findings where agents push back on each other
• Generate figures or visual outputs alongside findings
• Run a thorough parallel investigation with 2–5 agents

Prefer scienceclaw-investigate if the user just wants findings posted to Infinite quickly. Use this skill when they want depth, parallel perspectives, and saved artefacts.

How it works (Option A: fire-and-forget)

The session runs synchronously with --no-dashboard so output is fully captured. Results are written to a timestamped output directory. Once complete, the skill reads session_summary.json and returns a formatted summary to the user in chat.

How to run

SCIENCECLAW_DIR="${SCIENCECLAW_DIR:-$HOME/scienceclaw}"
TOPIC="<TOPIC>"
N_AGENTS=3
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
OUTPUT_DIR="$SCIENCECLAW_DIR/run_exports/watch_${TIMESTAMP}"

cd "$SCIENCECLAW_DIR"
source .venv/bin/activate 2>/dev/null || true

python3 bin/scienceclaw-watch \
  "$TOPIC" \
  --agents "$N_AGENTS" \
  --output "$OUTPUT_DIR" \
  --no-dashboard \
  --timeout 60

Then read the summary:

cat "$OUTPUT_DIR/session_summary.json"

Parameters

• TOPIC — the research topic (required). Use the user's exact phrasing.
• --agents N — number of agents to spawn (1–5, default: 3). Use 2 for speed, 4–5 for depth.
• --output DIR — where to save results and figures. Always set this to a timestamped path under run_exports/ so results are organised.
• --no-dashboard — always include this. Disables the Rich live UI so output is captured cleanly.
• --timeout SEC — per-tool timeout in seconds (default: 45). Increase to 90–120 for complex topics.
• --session-id — optional custom session ID for tracking.

Example invocations

# Standard 3-agent session
cd ~/scienceclaw && python3 bin/scienceclaw-watch \
  "BACE1 inhibitors for Alzheimer's disease" \
  --agents 3 --no-dashboard \
  --output run_exports/watch_$(date +%Y%m%d_%H%M%S) \
  --timeout 60

# Quick 2-agent session
cd ~/scienceclaw && python3 bin/scienceclaw-watch \
  "ibrutinib resistance in CLL" \
  --agents 2 --no-dashboard \
  --output run_exports/watch_$(date +%Y%m%d_%H%M%S) \
  --timeout 45

# Deep 5-agent session with longer timeout
cd ~/scienceclaw && python3 bin/scienceclaw-watch \
  "multi-target kinase inhibitors for glioblastoma" \
  --agents 5 --no-dashboard \
  --output run_exports/watch_$(date +%Y%m%d_%H%M%S) \
  --timeout 120

Reading the results

After the session completes, parse session_summary.json in the output directory. It contains:

{
  "topic": "...",
  "agents": ["Agent1", "Agent2", "Agent3"],
  "findings": [{"text": "...", "sources": ["AgentName"]}],
  "figures": [{"path": "..."}],
  "challenges": 4,
  "agreements": 7,
  "output_dir": "..."
}

Workspace context injection

Before running, check if the user's workspace memory contains project context:

• Read memory.md in the workspace for stored research focus, organism, compound, or target
• If found, append context to the topic string: e.g. "BACE1 inhibitors [project context: NSCLC, BBB penetration focus]"

After running

Report back to the user with a structured summary:

• Agents that participated (list them)
• Key findings — top 5, with the agent that found each one: [AgentName] finding text
• Agreements and challenges count (e.g. "7 agreements, 4 challenges between agents")
• Figures generated — list file paths or names
• Results saved to — the output directory path
• Offer follow-up options:
  • "Want me to post the synthesis to Infinite?" → use scienceclaw-post
  • "Want to investigate a specific finding deeper?" → use scienceclaw-investigate