Agentic Discovery Methodology

Every engagement runs with a small roster of LLM agents wired into the work. They aren't the deliverable. They're how a one-operator studio sustains the throughput of a fifteen-person agency without losing the judgment of one person.

What we mean by agentic discovery

Discovery is the part of an engagement where the studio learns the client's domain well enough to make competent recommendations. Traditional consultancy does it with interviews, document review, and one billable senior reading the corpus for two weeks. We do it with interviews, document review, and a small roster of agents reading the corpus alongside the senior.

The agents don't replace the reading. They make the reading auditable. Every claim the studio later makes about the domain is backed by a queryable trace: which document, which passage, which embedding cluster, which extraction pass.

The roster

Five agents do most of the work. The names are descriptive; the prompts are what differs from off-the-shelf wrappers.

• Corpus ingestion agent. Takes the client's document corpus (briefs, transcripts, prior research, internal wikis). Cleans, chunks, embeds, deduplicates. Output: a vector store and a chunk-level provenance map.
• Entity-extraction agent. Reads the chunks, proposes a domain ontology, extracts entities and relationships into a graph store. Output: the seed of the GraphRAG layer (see GraphRAG for Client Domains).
• Interview synthesis agent. Takes raw interview transcripts and produces structured findings tied back to specific quotes. Output: a findings document where every claim links to its source.
• Statistical-honesty agent. Reads every quantitative claim the studio is about to publish and grades it against the four-tier measurement policy. Refuses claims that can't earn a tier.
• QA agent. Runs the final pre-publish checklist: tier tags present, banned vocabulary absent, anti-AI-tells caught, internal-only references redacted.

The statistical-honesty agent in detail

This is the agent that does the work the studio is most likely to skip if it were running on humans alone. The temptation to inflate a number is real and constant. The agent doesn't get tired of asking the same four questions: where did the number come from, what was the sample size, was there a control or baseline, who counted.

If the answers support VALIDATED, the agent passes the claim with the green tag. If they support DIRECTIONAL, it downgrades. If the number's source is public or client-supplied, REPORTED. If there's a story but no number, ANECDOTE. If the answers don't support any tier, the claim doesn't ship and the agent says so in writing.

The agent overrules the studio when the studio is wrong. That's the entire point. An agent that always agrees is an expensive autocomplete.

The QA agent's checklist

The QA agent runs every page before it's published. The checklist is short by design.

1. Every quantitative claim has a tier tag adjacent to it.
2. No banned vocabulary (the studio keeps a list; the agent updates it when new tells emerge).
3. No em-dashes (commas, periods, parens instead).
4. No openers we've banned (the standard list of AI-tell phrases).
5. No parallel triplets to close sections.
6. No internal tool names or NDA-protected terms (the agent maintains a redaction list).
7. No client-confidential data outside the agreed disclosure scope.
8. Internal links resolve. External links return 200.
9. Image alt text present.
10. The page passes a final read-aloud check: does it sound like a person who knows the work, or like a generic SaaS company.

The list isn't sophisticated. Most quality problems are caught by a small set of repeatable checks run consistently.

When agents help

Three places agents earn their compute reliably.

Surface area larger than one person can read. A client with sixty subjects worth of content, or fifteen years of recorded sessions, or two thousand support tickets. The studio's senior can read enough to build judgment. The agents read everything, so the studio's recommendations are backed by the full corpus, not the slice the human had time for.

Consistency across long documents. Voice, terminology, claim discipline, accessibility patterns. Humans drift across a forty-page document. Agents don't.

Iteration speed on creative variations. Five variants of a script section, ten visual concepts for a moment, twelve ways to frame the same finding for different audiences. The studio still picks. The agent makes picking from twelve as cheap as picking from three.

When agents don't help

Three places we've watched agentic discovery fail or backfire.

Taste calls. Which of these is better is not a question we delegate. The studio decides. The agent's job is to surface the options, not pick among them. Outsourced taste reads as outsourced taste, and the engagement loses the thing the client hired us for.

Client-relationship moments. Hard conversations, scope conflicts, sign-off pushes, news the client doesn't want to hear. The agent can draft, but the human sends, with the human's name attached and the human's judgment intact. Mis-using agents here destroys trust permanently.

Domains where the corpus is misleading. When a client's internal documents systematically misrepresent their actual practice (which happens more than people admit), agents trained on the corpus amplify the misrepresentation. The fix is interviews, observation, and old-fashioned skepticism from the human senior. Agents make this failure mode worse, not better.

The public demo

The live intake agent on this site is the smallest possible version of the methodology. It runs a discovery conversation, classifies the inquiry against the four service tiers, writes a structured brief, and hands the brief to the studio for review. It's not the agent stack a Constellation engagement uses. It's the same shape at a fraction of the scope: discover, structure, hand to a human, audit.

Anything more ambitious than the intake agent runs inside an engagement, behind an NDA, with the client's corpus on the studio's instance. The intake is the only one that gets to be public, because it's the one operating on inputs the public is giving us.

What this isn't

Not an autonomous agent platform. Not a chat product. Not a vendor reselling a frontier model with a wrapper. The agents are tooling the studio runs to do its own work better. They show up in deliverables as captured methodology and audit trails, not as the deliverable itself.

Clients buy the work. The agents are how the studio makes the work hold up.