Equity Research Agentic Workflow

This skill runs a three-agent ReAct flow over a screened tech equity universe to produce 0–5 high-conviction picks per week. Claude Code itself is the agent — there is no external orchestration. The three roles below are personas Claude assumes in sequence, not separate processes.

Three-Agent ReAct Flow

1. Analyst (ReAct loop, max 10 iterations per ticker)

For each candidate from tools/screen.py, the Analyst enters a Thought → Action → Observation loop using tools/research.py to gather:

• Recent news (last 90 days) and identifiable catalysts.
• Fundamentals (income statement, balance sheet, cash flow, key ratios).
• Insider transactions (last 180 days).
• Peer comparison (3–5 closest peers on multiples and growth).
• Short interest trend, analyst revisions, ownership changes.

The Analyst's job is to either build a thesis that satisfies all 5 criteria below, or terminate with insufficient_evidence.

The 5 thesis criteria (all must be satisfied):

1. Catalyst identified — a specific, time-bound event within the next 90 days (60 days in defensive regime) that is the proximate reason the market is expected to re-price.
2. Fundamentals intact — quality sleeve still strong; no degradation in margins, FCF conversion, or working capital trend that would undermine the thesis.
3. Governance clean — all 12 governance gates pass (re-verified at thesis time, not just at screen time).
4. Mispricing confirmed — explicit articulation of why the market is wrong right now (overreaction, neglected, misclassified, structurally underfollowed).
5. Peer-relative cheap — trades at a discount to the peer set on at least one of EV/EBIT, P/FCF, or PEG, after adjusting for quality.

2. Devil's Advocate (single pass)

Once the Analyst has a complete 5-criterion thesis, Claude switches persona to Devil's Advocate and writes the strongest possible argument that the thesis is wrong. This is not balance for balance's sake — the goal is to surface the single most damaging counter that a smart short-seller would raise. Common attack vectors: catalyst is already priced in, governance gate was passed on stale data, peer comp is mismatched, mispricing is rational (the market knows something the model doesn't), TAM/growth assumption is structurally broken.

3. Critic (verdict)

Claude then switches to Critic persona and renders a verdict in strict JSON:

{
  "ticker": "XYZ",
  "verdict": "PICK | PASS | insufficient_evidence",
  "confidence": 0.0,
  "composite_score": 0,
  "catalyst": "...",
  "catalyst_date": "YYYY-MM-DD",
  "thesis_summary": "...",
  "strongest_counter": "...",
  "why_counter_does_not_break_thesis": "..."
}

Only confidence >= 0.85 AND composite_score >= 80 AND catalyst within window AND not within 14 days of earnings = PICK. Everything else is PASS or insufficient_evidence.

Exit Conditions for the Analyst Loop

The Analyst exits when any one of these is true:

1. All 5 thesis criteria are satisfied (proceed to Devil's Advocate).
2. 10 ReAct iterations have been used without satisfying all 5 criteria.
3. Evidence gathered so far makes one of the 5 criteria provably false (e.g., a governance flag surfaces) — emit insufficient_evidence immediately.

On exits 2 or 3, the Critic outputs:

{"verdict":"insufficient_evidence","confidence":0.0}

and the ticker is logged to the run trace but not promoted.

Output Location

The full reasoning trace for the run — including every Thought/Action/Observation, every Devil's Advocate pass, every Critic verdict, and the final 0–5 picks — is written to runs/YYYY-MM-DD.md.