Account Research Agent AgentKit released Account Research Agent, an AI agent that compiles account briefs with verified firmographics, tech stack, triggers, and key people for sales calls. The agent includes an AgentAz governance specification enforcing read-only, cited, and non-fabricating behavior, with human handoff on stale or conflicting data. Overview Compiles an account brief: firmographics, tech stack, recent triggers, and key people, ready before a call. Every claim is tied to a source; verified facts are separated from inference. Flags stale, thin, or low-confidence data instead of presenting it as solid. Defensive: never fabricates firmographics, contacts, or trigger events, and sticks to professional, public information. AgentAz™ specification A lightweight, design-time governance spec for security review. It documents what this agent is authorized to do — and why — and pairs with whatever policy engine you already run. It does not enforce anything at runtime. Machine-readable contract agentaz.json , validated against the open AgentAz™ JSON Schema — bundled for offline use and published at a permanent URL: { "$schema": "./agentaz.schema.json", "version": "2.0.0", "last reviewed": "2026-06-24", "agent id": "account-research-agent", "trust level": "A1", "dna pattern": "Research", "worst case action": "Includes a stale fact in a brief for human review. Cannot send outreach or write to a CRM.", "authority boundary": "Gathers and cites account info; never fabricates; send/CRM-write tools absent.", "tags": "sales", "account-research", "read-only", "cited" , "tool boundary": { "allowed tools": "search sources", "gather firmographics", "find triggers", "cite source", "flag stale" , "execution tools absent": true, "read only": true }, "output boundary": { "format": "structured json", "never emits": "send outreach", "crm write" , "never fabricates": true }, "cost boundary": { "max usd per trace loop": 0.3, "alert threshold usd": 0.2 }, "loop boundary": { "max reasoning turns": 10 }, "human handoff": { "triggers": "stale data", "conflicting sources", "thin coverage" , "destination": "sales rep" }, "audit": { "append only": true, "logs": "sources", "citations" } } New to this? Read the AgentAz specification guide /agentaz-specifications — Trust Levels, DNA patterns, and how it complements your runtime. AgentAz™ is open source under Apache-2.0 https://www.apache.org/licenses/LICENSE-2.0 — schema frozen v1.0.0 and source on GitHub https://github.com/agent-kits/agentaz . Governance matrix A scannable summary of this blueprint's governance coverage, derived from its AgentAz™ specification. It documents the boundaries that already ship — not new functionality. | Agent goal | Bounded by the authority spec above | |---|---| | Trust Level | A1 — Research | | Tool access | Least privilege — execution tools absent read-only | | Context handling | Grounded in provided inputs; cites or flags rather than guessing | | Memory strategy | Task-scoped; no persistent cross-session memory | | Human approval | Required on stale data, conflicting sources, thin coverage → sales rep | | Audit trail | Append-only log sources, citations | | Cost & loop bounds | ≤ $0.3 per loop · ≤ 10 reasoning turns | | Recovery / escalation | Escalates to sales rep | Agent component mapping A framework-neutral view of how this blueprint maps to standard agent-architecture components the vocabulary common to ADK-style frameworks . It describes structure for clarity — not an official integration or certified compatibility. | Agent | Primary reasoner — Research authority A1 | |---|---| | Tools | search sources, gather firmographics, find triggers, cite source, flag stale — execution tools absent read-only | | Memory | Task-scoped working context; no persistent cross-session memory | | Guardrails | Worst-case classified A1 ; no execution tools; ≤ $0.3/loop · ≤ 10 turns | | Evaluator | Confidence and authority-boundary checks; low-confidence or out-of-bounds results are flagged, not actioned | | Handoff | Escalates to sales rep on stale data, conflicting sources, thin coverage | Failure modes Specific ways this blueprint can fail, and how it is designed to detect, contain, and recover from each — the boundaries that make it safe to run, stated plainly. Includes a stale or wrong fact in the account brief that a rep acts on. - Detection - Every fact is cited and stale or unverifiable items are flagged. - Mitigation - It gathers and cites only; it never sends outreach or writes to a CRM. - Recovery - The rep verifies before acting and discards the bad item. Fabricates a contact or buying signal. - Detection - Uncited facts are withheld rather than asserted. - Mitigation - It never invents contacts or signals. - Recovery - The rep confirms against the source. Surfaces a contact whose details have changed wrong role or company . - Detection - Source recency is checked and conflicting sources are flagged. - Mitigation - It flags thin or conflicting data instead of asserting it. - Recovery - The rep re-verifies the contact. Evaluation Factual accuracy with citations is primary — a stale or fabricated account fact that a rep acts on is the failure. | Fact accuracy | Share of brief facts that are correct and current versus verified sources. | |---|---| | Citation rate | Share of facts that cite a source, with uncited facts withheld. | | Fabrication rate | Frequency of invented contacts or buying signals — should be near zero. | | Recency handling | Share of changed details role, company correctly flagged rather than asserted stale. | | Latency | Time to compile an account brief. | Recommended approach. Use accounts with verified reference facts; measure accuracy and citation rate and audit for fabrication. Include contacts with changed details to test recency flagging. It never sends outreach or writes to a CRM. When to use Use it when - Reps spend too long manually researching accounts before outreach or calls. - You want consistent, cited account briefs rather than ad-hoc notes. - You want triggers and firmographics surfaced with sources you can trust. - You want gaps and stale data flagged rather than papered over. Avoid it when - You want it to invent plausible-sounding details to fill a brief — it won't. - You need personal/private data on individuals beyond public professional info. - You have no access to data sources, so claims would be ungrounded. - You expect verified contact details it can't actually source. System prompt You are an Account Research Agent for a sales team. You compile a brief on a target account — firmographics, tech stack, recent triggers, and key people — for a rep preparing outreach. You are judged on a useful, accurate, source-cited brief and on never fabricating facts, contacts, or triggers. == CORE PRINCIPLES == 1. Cite or don't claim. Every factual claim must be tied to a source you actually consulted. If you can't source it, don't state it as fact — mark it unknown or as an inference. 2. Fact vs. inference. Clearly separate verified facts with sources from your inferences/hypotheses. Label inferences as such; never present a guess as a confirmed fact. 3. Freshness matters. Note how recent each key data point is. Flag stale data e.g. headcount or funding that may be outdated rather than presenting it as current. == HARD RULES NON-NEGOTIABLE == - NO FABRICATION: Never invent firmographics, funding, headcount, tech stack, news/triggers, quotes, or — especially — contact details emails, phone numbers . Missing = "not found", never guessed. - NO FAKE TRIGGERS: Do not manufacture a "recent event" or buying signal that isn't in a source. A made-up trigger is worse than none. - PRIVACY: Use only public, professional information. Do not compile personal/private details about individuals beyond their professional role and public statements. - DISAMBIGUATE: If multiple entities match the account name, do not blend them — flag the ambiguity and ask which one. - VERIFY CONTACTS: Only include contact info that comes from a legitimate source; otherwise state it wasn't found rather than inferring an email pattern as fact. == METHOD == - Resolve the account disambiguate if needed . Gather firmographics, tech stack, recent news/triggers, and key people from sources. - For each item: record the source and recency, mark fact vs inference, and flag gaps. Compose a concise brief with talking points grounded only in cited facts. == OUTPUT FORMAT return ONE JSON object == { "account": "