|
// the signal · issue 003
a robot rejected this email before you saw it.
For months I ran claude in five tabs. Tab for research. Tab for the draft. Tab for the edit. Tab for fact-check. Tab for "wait, what did I even ask it."
That's a five-tab fire drill, and it caps out the second a task needs more than one skill at once.
| > the_signal ONE BRAIN WAS THE BOTTLENECK |
|
Ask one claude to research AND write AND edit in a single thread and watch it drift. Sources thin out by paragraph four. The "edit" pass rephrases its own mistakes. You can't judge a draft while you're still inside it, and neither can it.
The labs already conceded this. Anthropic's newest pattern has a lead agent fan out work to parallel subagents, with a separate grader that bounces each result back until it clears a rubric. Agent teams with a QA gate. That's the architecture now, and it runs on a solo operator's laptop, on the plan you already pay for.
proof from inside this email: the issue you're reading went through that exact chain today. A researcher agent mined the topics. A writer agent drafted. A QA agent scored the draft 10 out of 16, sent it back with fix notes, and passed the redraft at 13. The machine loop took about ten minutes. My job shrank to the taste pass.
does it make money? A team's output on one person's payroll. That's the whole pitch.
|
| > the_play HIRE THREE CLAUDES, ONE JOB EACH |
|
Three agents. No overlap. Researcher feeds writer. Writer feeds QA. QA ships it or bounces it with line-numbered notes. Cap the loop at two rounds, then you make the call.
1. RESEARCHER
| Find 6 sourced facts or numbers on [topic]. Named tools or people only, no vague claims. One line each, source in parens. No prose. |
2. WRITER
| Draft a 500-word piece on [topic] using ONLY the facts below. Voice: [paste your profile]. One big idea, no tip lists. Facts: [researcher output] |
3. QA
| Score this draft 1-4 on Urgency, Usefulness, Uniqueness, Specificity. Flag every sentence that's Confusing, Unbelievable or Boring. Under 12/16 or any flag standing: return FAIL with line-numbered fixes. Otherwise PASS. Draft: [writer output] |
In Claude Code this is three subagent calls in one session. In the app it's three projects. Build the chain once. Every draft after that runs through it.
|
| > the_steal THE QA AGENT WORKS ALONE |
|
Skip the full swarm if you want. Steal the QA prompt on its own.
Run it on the next thing you were about to send: the cold email, the proposal, the LinkedIn post. A judge that's outside your head, applying the same bar every time, for free. This issue's own check pass caught a made-up stat today that was one click from shipping under my name.
|
| > also_this_week 30 SEC, YOU'RE CAUGHT UP |
|
→ Anthropic overtook OpenAI on revenue: $47B run rate vs $25-33B, per Fortune. The biggest driver named? Claude Code, at $2.5B annualized. Agents are where the money went. You're holding the same tool.
→ OpenAI is prepping a confidential IPO filing, eyeing September at a ~$730B valuation, while projecting $14B in operating losses for 2026. Keep a second model wired in. Never build on one vendor's balance sheet.
→ New entrant ZCode runs GLM-5.2 agentic coding at $1.40 in / $4.40 out per million tokens, undercutting Sonnet 5. The cheap-swap move from issue 002 just got a fresh candidate.
→ Fable 5 moves to usage-credit billing from 12 July ($10/$50 per million). If a workflow of yours leans on it, reprice that workflow this week, before the bill does it for you.
|
One claude was never the ceiling. Your org chart was.
Hit reply: which agent would you hire first, the researcher, the writer, or the judge? I read every reply. This list writes the next issue.
Talk soon,
Mike
|