← All workflows
Research

Market research synthesis (agentic loop)

Multi-source research agent that autonomously browses, retrieves, and synthesizes a structured sector brief.

Agentic sector brief on a PitchBook + Box + Slack stack: $5.51/run, 51% invisible. Firedog maps every iteration hop, accumulation spike, and connector pull.

Model 26%Prompt 23%Orchestration 11%Retrieval 4%History 25%Connectors 11%
ComponentWhy it is unattributedcalls$ / run
Model · 26%LLM inference (Claude Opus 4)Each agentic step sends the current research state plus new retrieved content to Opus 4. Iteration count varies 8-22 steps per brief.tokens in: 55,000 · tokens out: 8,000Visible on the invoice.14$1.43
Prompt · 23%Research persona and output schema prefixA 6,000-token system prompt defining research depth, output sections, and citation style is resent on every iteration without prompt caching enabled.tokens in: 84,000 · tokens out: 0Visible on the invoice.14$1.26
Orchestration · 11%Planner-executor-verifier loop hopsEach iteration invokes a planner subagent, one or more executor agents, and a verifier; each is a separate billed call. A 14-step brief produces 42 billed hops.tokens in: 18,000 · tokens out: 2,200Each planner-executor-verifier trio looks like one 'research step' in logs; the three underlying API calls are never surfaced as separate line items.42$0.60
Retrieval · 4%Iterative web and vector retrievalThe agent fetches and re-embeds sources on each loop iteration. Retrieval is not deduplicated across steps; overlapping sources are re-embedded.tokens in: 22,000 · tokens out: 0Retrieval cost sits on the vector-DB vendor invoice; no cross-charging to the workflow P&L occurs.28$0.24
History · 25%Full research thread re-sent each iterationThe complete running research memo is prepended to every iteration call so the agent retains prior reasoning. By step 14 this is 60k+ accumulated tokens sent each time.tokens in: 140,000 · tokens out: 0Accumulated research state is framed as 'context' in framework logs, not as re-billed tokens, so it does not appear in per-call spend reporting.14$1.36
Connectors · 11%PitchBook, Box, Google Workspace, and Slack connectorsEach brief calls PitchBook for market data, reads/writes source exhibits in Box, exports the final memo to Google Docs, and posts to Slack. Pulled records are re-normalized by an LLM step so external fields fit the research schema; this repeats per iteration because retrieval is not deduplicated.tokens in: 20,000 · tokens out: 1,500PitchBook API usage sits on the data-vendor contract and the Box/Google/Slack connector calls on the iPaaS bill; none of it is attributed to the agent's per-brief cost.12$0.62
One agentic sector brief costs ~$5.51 to run; a 10-analyst fund burns ~$82.7k/yr and 51% is orchestration overhead, re-billed accumulation, and PitchBook/Box/Slack connector sync no dashboard shows.
$5.51
cost per run
$83k
per year
51%
unattributed today
$33.08
per user, per day
Scenario: 10-analyst macro fund, weekly cadence. 10 active users, 6 runs/day, 250 working days. 30 documents/run. Avg doc: 5,000 tokens.

Default model: claude-opus-4

Trigger: Weekly sector updates plus on-demand for event-driven trades; 3-8 briefs per analyst per week.

Who runs it: Macro and sector companies, VC research teams, research pods. A single brief typically drives 15-40 minutes of wall-clock agentic runtime. Briefs are grounded in PitchBook market data, sourced and archived in Box, exported to Google Workspace, and delivered in Slack, so each iteration touches that connected stack.

What Firedog shows

  • Token cost per agentic iteration, showing the inflection point where history accumulation overtakes inference spend.
  • Which executor agent calls are producing duplicate retrieval and how much that costs per brief.
  • Projected $/year reduction from capping iteration depth at 10 steps with a summary-compression step.
  • How much of each brief is PitchBook data pulls and connector normalization repeated across un-deduplicated iterations.

Decisions this informs

  • Implement a rolling summary compression after step 5 to prevent unbounded history accumulation.
  • Deduplicate retrieved sources before re-embedding; roughly 30% of retrieval calls hit already-ingested content.
  • Route planner and verifier hops to Haiku 4; run only executor calls on Opus 4.
  • Cache PitchBook pulls and Box reads for the life of a brief so the same records aren't re-fetched and re-normalized on every agent iteration.

Connected systems

  • PitchBook · Market data · Morningstar (PitchBook)
    The agent pulls private-market comps, financings, and company profiles from PitchBook on each brief and normalizes them into the research state.
  • Box · Document storage
    Prior briefs and source exhibits are read from and archived to Box.
  • Google Workspace · Productivity · Google
    The finished sector brief is exported to Google Docs / Drive for the analyst's team.
  • Slack · Collaboration · Salesforce (Slack)
    Completed briefs and mid-run status are delivered to the desk's Slack channel.

Understand every AI dollar.

Book a demo to see how Firedog attributes every AI request across your organization, and what your leadership team can do with the data.

Book a demoBecome a design partner