Directive to deliverable: watch one sentence turn into a finished investor memo
The whole premise of NEXUS is that you give a directive and get a deliverable. Not a prompt and an output. Not a configuration and a pipeline. A directive in plain language, and a deliverable in final form.
This post is the walkthrough of what happens between those two things, one real example, end to end, with every specialist that ran and what it did.
The directive
"Write an investor memo for a $1.5M seed round. Company is a B2B AI orchestration SaaS for mid-market teams. 3-5 pages. Cover: problem, solution, market, traction, team, ask."
That's what goes in. Plain English. No template, no fill-in-the-blanks, no magic words. The orchestrator receives it, parses it, and starts assembling a team.
Seconds 0-3: the orchestrator decomposes the directive
First it works out what the deliverable actually requires. An investor memo is a well-understood format: problem, solution, market, traction, team, ask. Length target: 3-5 pages, roughly 1,500-2,500 words. Audience: institutional investors, so the tone is confident, data-driven, no fluff. It produces an internal plan: eight specialists, parallelism where possible, a council debate on tone and positioning, and a PM agent to keep the narrative consistent across sections. None of this is shown to you. You just see "working…"
Seconds 3-20: research and structure, in parallel
Three specialists dispatch at once:
- Market-research agent (Gemini Pro), current B2B SaaS market data, AI-orchestration TAM estimates, comparable valuations from recent rounds
- Competitive-landscape agent (Claude Opus), maps the orchestration competitive set and NEXUS's differentiation angles (BYOK billing, the 100-agent architecture, shared memory)
- Structural-outline agent (Claude Opus), designs the six-section memo with word budgets, key points, and a narrative arc
It's a fan-out. Total elapsed: 17 seconds, the slowest of the three.
Seconds 20-55: the section writers dispatch
With research and structure in hand, six writing specialists go out, one per section:
- Problem writer (GPT-4), a punchy account of why mid-market teams struggle with AI
- Solution writer (Claude Opus), the NEXUS architecture, pitched to non-technical investors
- Market writer (GPT-4), TAM/SAM/SOM, built from the Stage 1 research
- Traction writer (Claude Opus), waitlist growth, conversion signals, early customer anecdotes
- Team writer (GPT-4), founder background, advisors, hiring plan
- Ask writer (Claude Opus), the funding ask, use of funds, milestones to the next round
Five run concurrently; the Ask writer runs last, because it references numbers from the Traction and Market sections. Each specialist has a tight system prompt, 800 to 1,200 words on nothing but how to write its section for this exact audience, and sees only its slice of the outline, not the others' drafts. Total elapsed: 33 seconds (five in parallel plus one serial tail).
Seconds 55-75: the PM agent runs a consistency pass
This is the step that separates orchestrated work from a naive multi-agent pipeline. The PM agent (Claude Opus) takes all six drafted sections and checks them against each other:
- Does the problem section's framing match what the solution section solves?
- Does the market section's TAM match what the ask section assumes for growth?
- Does the traction section's conversion math support the ask's valuation?
- Is the tone consistent, or did one specialist drift casual while another stayed formal?
Two flags come back. The market section cites a $45B TAM; the ask section implies $60B through its growth assumptions. The PM catches the contradiction, and the orchestrator triggers a correction agent (Claude Opus) to reconcile the numbers across both. The second flag: the problem section says "teams" where the solution section says "organizations." Minor, but flagged. A quick pass standardizes the terminology.
This is the jump from roughly 40%-correct to 95%-correct that most multi-agent systems skip. Without the PM agent, both flaws ship, and the memo looks professional at a glance, then falls apart under an investor's actual scrutiny. Total elapsed in this stage: 16 seconds.
Seconds 75-85: council debate on positioning
One more specialist runs, a positioning specialist (Claude Opus), which reviews the full memo and makes a single judgment call: is the differentiation sharp enough? Is NEXUS clearly framed against the alternatives, or does the reader have to guess why it beats AutoGen, CrewAI, or LangGraph?
It decides the framing is too soft: the BYOK pricing advantage should be explicit, not implied. That's a judgment call, so the orchestrator escalates to the council. Three specialists convene: the solution writer, the ask writer, and a positioning strategist. They debate whether naming competitors in an investor memo reads as bold or risky. Verdict in 8 seconds: name them. Investors already know the competitive set; hiding from it reads as weakness. The orchestrator applies the verdict, and a small revision adds the named comparison to the solution section. Total elapsed: 10 seconds.
Seconds 85-90: final polish and assembly
A grammar specialist (local Llama on owned servers, free) runs a last mechanical pass, awkward phrasings fixed, passive voice trimmed, em-dashes standardized. The orchestrator assembles the final output: a 2,100-word investor memo, six sections, consistent terminology, reconciled numbers, sharp competitive positioning, no grammatical issues. It lands in your interface as a downloadable document.
What just happened
Eight specialists. Four providers. About 90 seconds of wall-clock time. Roughly $0.94 in API fees.
And the part that matters: you did none of the orchestration. You wrote one sentence. You didn't pick models, design the pipeline, write a system prompt, run the consistency pass, or escalate the positioning question. You gave a directive; the system produced a deliverable. That is the entire product.
What this is not
To be clear, this isn't a claim that the generated memo beats a seasoned operator's hand-written one. A founder with years of pattern-matching on investor psychology, real relationships, and specific market insight will produce something sharper.
It is a claim that this memo beats a first-time founder staring at a blank document at 11pm, googling "investor memo template," stitching together examples from YC archives, and producing something stiff by 3am. The 90-second draft is a strong starting point. The founder should still read it, push back, rewrite the parts that feel wrong, and add the specific anecdotes only they can. Half an hour of human polish on an orchestrated draft beats five hours of solo writing from scratch. That's what orchestration is for, not replacement, leverage.
The whole thing, in one frame
Directive in, deliverable out. Eight specialists, four providers, about 90 seconds, under $1 in API cost. No prompt engineering, no tool configuration, no pipeline design, no model selection, no system prompts to write.
That's the product, and it's why we built it. If that sounds like something you want for your own work, research, writing, analysis, outreach, content at scale, join the waitlist. We launch in about 60 days.
That wraps the first editorial arc. Next come deeper dives into specific use cases, customer tutorials once we launch, and the occasional behind-the-scenes note on the build. Stick around if you want to watch this thing ship.