Putting Nine Research Specialists Behind One Question: A Multi-Agent AI System for Pare & Co
Scout runs account, partner, vertical and competitive intelligence for the leadership team from Slack. Nine specialist agents do the research; Scout’s job is to know what to ask and to make sure the answer arrives.
The work
The Challenge: The People With the Context Do Not Have the Afternoon
A consultancy sells judgment about markets it has to understand faster than its clients do, and the intelligence work behind that judgment is unbounded: which health systems just hired a chief digital officer, which accessibility deadline is about to force a remediation budget, where a competitor is publishing something the firm should have written. Doing it properly for one account means thirty or more searches, cross-referencing hiring pages against press releases against the CRM against a technology fingerprint of the site.
No one with a P&L has an afternoon for that, and the people who do have the afternoon do not have the context to know which finding matters. So the work got done occasionally, for whichever account had someone’s attention that week — and every account deserves it every week.
The Strategy: A Multi-Agent System on the Orchestrator-Worker Pattern
Scout is a multi-agent AI research system built on the orchestrator-worker pattern: one orchestrator that researches nothing, and nine specialists that do. The specialists are built in Claude Managed Agents, each a standing configuration with its own system prompt, tools and output contract.
Leadership asks in Slack, in plain language. Scout works out which specialist the question needs, assembles the dispatch and returns a sourced brief to the person who asked.
- Context engineering is the orchestrator’s job. Scout holds the internal context: relationship owner, prior brief, CRM state, which business unit the answer is for. It is forbidden from telling a specialist how to format its answer, because an orchestrator that dictates structure loses everything the specialist would have surfaced unprompted.
- Delivery is engineered with the same care as the research. Agents are polled to completion, every task and session id is written to memory at dispatch, results move to durable storage the moment they exist and every run ends with a message back to the requester.
- Guardrails, with a human in the loop. No code, no email, no calendars, no projects. Scout creates the plan and the draft; a person moves them to action.
- Agent memory builds on itself. Continuity is files in version control: daily notes, a curated long-term record and standing instructions. A correction given once changes every draft after it.
The Outcomes: A Standing Research Function, Asked in Slack
The intelligence work happens now. A question in Slack is enough to start it, where a free afternoon used to be the price.
In production since April 2026, Scout answers with briefs the team can read, cite and act on: account intelligence before an outreach, a vertical sweep before a pursuit, a competitive read before a proposal. Business development runs it as a standing research function, and the leadership team asks it the questions that used to wait for whoever had time.
And because the memory builds on itself, the system keeps getting more useful to the organization that runs it. The judgment stays human; what changes is what that judgment rests on — current intelligence about every account, partner and market the firm is responsible for.
The detail
Why it mattered
- The work is real and it is unbounded. Which health systems just hired a chief digital officer. Which accessibility deadline is about to force a remediation budget. Which partner program the firm is quietly falling behind in. Where a competitor is publishing something we should have written. Doing it properly for one account means thirty or more searches, cross-referencing hiring pages against press releases against the CRM against a technology fingerprint of the site.
- The people with the context do not have the afternoon. No one with a P&L has half a day for that, and the people who do have the afternoon do not have the context to know which of the findings matters.
- A research tool answers the question you already knew to ask. The questions that actually arrive in Slack are each their own kind: someone from this company just hit our site, who are they and what do we say; do the same sweep you did for higher ed, but for healthcare; who built this university’s new site; we just lost a deal in this category, what does the rest of the category look like. Each has a different set of sources and a different definition of done.
- And the answer is worth nothing until it reaches the person who asked. In the form they can act on. Delivery is an engineering problem in its own right, and it is designed for with the same care as the research.
How it was built
- The orchestrator-worker pattern, in production. Scout runs on OpenClaw, persistent and Slack-native with a workspace on disk, and its instructions are explicit: dispatch the specialists, orchestrate, compare, summarize and advise. The pattern keeps one accountable brain on the question while nine specialists work it in parallel.
- Nine specialists, built in Claude Managed Agents. Account, Partner, Vertical, Competitive and Self-Intelligence, which reads what the market says about us, plus Content Compass, which judges whether a proposed article is worth writing, an Article Writer, a prospect technology fingerprinter, and a CRM agent. Each is a standing configuration with its own system prompt, tools and output contract. Six run on Opus at extra-high effort; the cheaper three take the mechanical jobs.
- They are not small. A single account brief consumes between 350,000 and a million input tokens and returns fifteen to twenty thousand: a half-day of desk research, compressed, with sources.
- Context engineering is the orchestrator’s real job. Scout holds the internal context, meaning who owns the relationship, what the prior brief said, what the CRM knows and which of four business units the answer is for, and assembles it into a rich dispatch. It is forbidden from telling the specialist how to format the answer, because an orchestrator that dictates structure gets the structure it asked for and loses everything the specialist would have surfaced unprompted.
- Voice is a loadable skill. The Article Writer has two: the firm’s house voice and the CEO’s personal register, the same two-layer model the humans write under, packaged so an agent loads it the way it loads any other tool.
- Guardrails, with a human in the loop. Scout does not code, does not send email, does not manage calendars, does not manage projects. It creates the plan and the draft; a person moves them to action. An agent doing business development is one bad afternoon away from emailing a prospect, and the cheapest place to stop that is in the definition of the job.
Built to deliver
- An answer only counts when it arrives. A dispatched agent is polled to completion inside the same turn. The task and its session id are written into memory at dispatch, so nothing is lost to an idle session. Results are pushed to durable storage the moment they exist. And no run ends without a message back to the requester, in the channel where they asked, including when a run fails: silence reads as broken.
- Reliability is what earns the second question. In an agentic system, the research is the easy half. The half that decides whether anyone trusts it is whether the work reliably arrives, in the place a human is already looking, on the day they asked. Scout’s delivery mechanics are designed around that principle.
Memory that builds on itself
- Continuity is files, and the files are in version control. Scout starts every session with no recollection of the last one, so its agent memory is explicit: daily notes hold raw events, a curated long-term file holds what turned out to matter, and a set of workspace documents function as standing instructions.
- A correction given once changes every draft after it. A managing director explained the outreach emails were too long; that went into the file and the drafts changed. The firm reorganized and a business unit’s name changed; the agent recorded the change and corrected its own model of the org chart.
Multi-agent AI system questions
What is the orchestrator-worker pattern?
One agent owns the question and other agents do the work. Scout receives the question, assembles the context, dispatches the right specialists and returns a sourced brief to the person who asked; nine specialist agents, each a standing configuration with its own system prompt, tools and output contract, do the research. The pattern keeps one accountable brain on the question while the work runs in parallel.
Should one AI agent do all the research itself?
No. Scout’s own instructions forbid it from doing deep research: it orchestrates, compares, summarizes and advises. A specialist with its own tools and output contract goes deeper on its slice than one generalist agent can go on everything.
Why do multi-agent AI systems fail in production?
Delivery, more often than research. In an agentic system the research is the easy half; the half that decides whether anyone trusts it is whether the work reliably arrives, in the place a person is already looking, on the day they asked. Scout is engineered around that: dispatched agents are polled to completion, every task and session id is written to memory at dispatch, results move to durable storage the moment they exist and every run ends with a message back to the requester, including a failed one.
How do AI agents remember between sessions?
Scout’s agent memory is explicit and lives in files under version control: daily notes hold raw events, a curated long-term record holds what turned out to matter and workspace documents function as standing instructions. A correction given once changes every draft after it.
What guardrails does an AI research agent need?
Written refusals and a human in the loop. Scout does not code, send email, manage calendars or manage projects. It creates the plan and the draft; a person moves them to action. The cheapest place to stop an agent from overstepping is in the definition of the job.
Client leadership
Christopher Murray
Founder & CEOChris Murray founded Pare & Co in 2007 and rebuilt the firm in 2026. He led Scout from concept to production, and is one of the leaders whose research it now runs.