Browse the reference ⌄
Worked Example published

Worked Example: Research Assistant

A composite Asset Manager X request is split into a grounded assistant and a deterministic routing workflow instead of a multi-agent research crew, then written down twice: as the ten-field design record and as a filled bounded-agency envelope, with two contrasting envelopes and their alert thresholds.

Version
0.2.0
Owner
DEIXEI
Last reviewed
Reading time
17 min

For Cto · Ai Architect · Product Leader · Platform Engineer · Risk And Assurance

Sources and details
Source works
Book 1 · Book 2 · Book 3
Source chapters
2, 22
Source appendices
02:C, 04:D
Refresh trigger
source change

Full source list ↓ Source index →

Case status. Asset Manager X, Bank A and Bank D are composites. This page shows decisions and records, not evidence about any real institution. Model identifiers, routes, costs and prices are the composites’ values as of May 2026, per Book 3 Appendix D, and will be out of date; treat them as the shape of the record, not as a recommendation of a model.

This page has three parts. Part 1 takes the product-shape decision (Book 2 Ch 22). Part 2 writes the result down as the ten-field design record and as a filled Claude envelope (Book 3 App D). Part 3 gives two contrasting envelopes (a Claude Code review subagent, and an HR assistant from Book 1) so you can see how the same four sections tighten or loosen with the job.

Part 1: the product-shape decision

The request

The research team receives about four hundred pages of broker research each morning. Analysts first decide which half-dozen documents matter, then compare them with the firm’s own position notes, holdings and committee minutes. The sponsor proposes a “research crew” of seven collaborating agents.

Apply the decision sequence

The full sequence and escalation criteria are on choosing the AI product shape. In the composite, it took fifty minutes.

QuestionAnswer in the caseConsequence
Is AI needed?Partly. Much of the morning volume is broker research arriving in known formats through two aggregators; deciding which documents to read is largely routing and deduplication, with no generation (the file-name distribution settled the argument)Cut out the part that needs no model
Prediction or generation?Both: relevance scoring is prediction; synthesis is generationTwo products with different owners, not one
Enterprise facts needed?Yes, decisively: the value is synthesis against the firm’s own notes, holdings and minutes, data only the firm holdsGrounded retrieval with permission-aware access and citations
Are the steps known?Yes: find relevant documents, read them against the firm’s position, note disagreements, draft a paragraphNothing needs to decide its own next step
Agency classClass 1 grounded assistant for synthesis; class 2 workflow in front for routing and deduplicationCode determines the next step in both
Reuse (tier by tier)Existing landing zone; the governed retrieval store another assistant already uses; model gateway with a mid-tier model and a small classifier already catalogued; workflow templates and connectors; conversation shell and citation componentSeven of eight capabilities inherited; two new ingestion connectors built
PolicyThe pre-approved policy pack for a class 1 grounded design over confidential data attached without a meetingSecond-line question resolved in minutes, not a six-week assessment
Hosting routePrivate tenancy, because the corpus carries confidential data and occasionally material non-public information (MNPI)The composite firm’s decision for its data classes, not a universal rule; map your own classes on data classes and lanes

Preserve the human boundary

The second line asked what stops an analyst treating a synthesised paragraph as research fit to send out. The answer went into the design record as a residual risk with a control: output carries its citations, the drafting surface is marked draft, and outbound compliance review is unchanged. The system assists analysts; it does not turn generated text into client-ready research.

What was rejected, and the revisit trigger

AlternativeDecisionReason
Class 2 only (workflow, no generation)RejectedSynthesis needs grounding against internal material
Class 3 bounded agent for peer comparison (find comparables, pull filings, build the table)Deferred with a triggerSteps are less predictable. Revisit when the evaluation harness holds a quarter of graded peer-comparison outputs
Class 4 multi-agent “crew”Rejected on the 3-to-4 evidence ruleSeven named roles, but no demonstrated single-agent degradation and every role reporting to the same head of research: no separate ownership boundaries

A deferral with a measurable trigger is a commitment, not a refusal. The “crew” was a metaphor, not a requirement.

Outcome (composite, read the following spring)

  • The routing workflow went live in five weeks and the grounded assistant in eleven, against a vendor’s eight-week promise for something roughly seven times larger.
  • Nine analysts use it daily. Usage telemetry showed the routing workflow, not the assistant, carried the larger share of time saved: unpredicted, and it changed what the next two requests looked like. Measure use before you claim value (factory measurement).
  • The two aggregator connectors were contributed back to the catalogue and reused by a regulatory-watch product.

Part 2: writing it down

The classification line

The six values the sequence produces, carried in the repository manifest and the catalogue entry (Book 2 Ch 22 §5.4; semantics on BOM and recipes):

reasoning and orchestration / knowledge and grounding / class 1 / tier 2 / confidential / composable capability

That is: anatomy tier / pattern family / agency class / risk tier / data class / reuse generation. Search, policy-as-code, the catalogue and the pattern arbiter all read it.

A six-value classification line joins anatomy tier, pattern family, agency class, risk tier, data class and reuse generation, read by catalogue search, policy-as-code and the pattern arbiter.

The classification line: one line, six values, carried in the repository manifest and repeated on the catalogue entry. Source: The AI Delivery Factory, Chapter 5, Figure 5.1.

The design record (ten fields)

An architectural decision record in a repository agents can reach, superseded rather than edited when it changes (Book 2 Ch 22 §5.5). Filled for this case:

#FieldThis case
1Problem and business ownerMorning triage and synthesis of broker research against internal positions; head of research
2Classification lineAs above
3Chosen agency class, and every lower class rejected with the reasonClass 1 assistant plus class 2 routing workflow; class 2-only rejected (synthesis needs grounding); class 3 deferred with trigger; class 4 rejected
4Patterns used per anatomy tierGrounded retrieval with citation enforcement and permission-aware retrieval; routing/deduplication workflow; conversation shell
5Catalogue assets reused; new assets proposedSeven of eight capabilities reused; two aggregator connectors proposed for contribution
6Model class and hosting routeMid-tier model and a small classifier via the gateway; private tenancy for confidential data and MNPI
7Policy pack and residual risksPre-approved class 1 confidential pack; residual: synthesis mistaken for publishable research (controls: citations, draft marking, unchanged compliance review)
8Evaluation plan and acceptance thresholdsGraded outputs in the evaluation harness; see the envelope below for the composite threshold
9Named accountable human for the operational scopeThe wrangler (see envelope)
10Exit or fallback strategyWhat happens if the model, vendor or pattern fails; see vendor concentration and exit

Sequencing rule (Book 2 Ch 22 §8): no branch until the record exists. A record committed after the first substantial pull request is retrospective justification; you can tell from the timestamps. If writing it takes more than an hour, the fields are wrong. Field 3 is the one that changes behaviour: a record that lists only the chosen class is a receipt; one that lists the rejected classes and why is a decision.

Research Paper 1 gives the same ten-field record (§7.5) with slightly different names: “classification code” for field 2, “processing lane” for the hosting route in field 6 and “control bundle” for the policy pack in field 7. This page uses Book 2’s names.

The envelope record: the Claude variant

The design record says what the product is. The envelope says what each running agent may do, with what, remembering what, and how it is watched. Book 3 Appendix D gives a filled record for Asset Manager X’s research-synthesis agent, envelope ASMX-RES-SYN-001, registered 14 March 2026 and reviewed at 60 days in production. The rule it enforces: no production Claude agent ships without a registered envelope. The runtime refuses to start without an envelope ID, the gateway refuses to forward without one in the headers, and a SOC rule fires on any successful model call that lacks one.

How the two books’ records relate. Book 2’s case and Book 3’s envelope both describe research synthesis at the same composite firm, written in different books, and they differ in two places. Book 2 chooses private tenancy; Book 3’s envelope routes through the direct Anthropic API behind the firm’s API gateway (Sweden Central with paired-region failover). Book 3’s agent also has WebFetch to five external financial-data domains, which Book 2’s design (internal corpus plus ingestion connectors) does not need. This page follows Book 2 for the product-shape decision and Book 3 for the envelope format and field values. Page recommendation: if you adopt Book 2’s design, remove WebFetch from Means unless the design record names a need for it, and set deployment_route to the route your data-class policy permits (deployment routes).

Mandate

FieldValueWhy
use_case_classresearch_synthesis
risk_tierT2Regulated process support, no autonomous customer impact
business_owner_name / _roleCéline, CIO
wrangler_name / _roleInes, senior research analyst, equity deskMust resolve to an active directory identity at registration and every review
named_approver_roleHead of investment research plus head of model risk (second line)Second line required for T2 and above
autonomy_ceilingdraftexecute_autonomously is allowed only at T0 and T1
fria_referencenullNot an Annex III use case; the trigger rules are on the regulatory map
usage_policy_alignmentNot investment advice; hitl_control_ref ASMX-HITL-001 (every synthesis signed by the analyst before external use); disclosure path research_note_footer[FACT-CHECK: the Anthropic Usage Policy version and high-risk category list Book 3 cites; confirm against the current policy]
customer_facingfalse
change_authority_refVirtual CAB minutes referenceRegistration is a change (change governance)

Means

FieldValueNote
deployment_routedirect_api_apimSee the route difference above
model_id / model_fallbackclaude-sonnet-4-6 / claude-opus-4-7 (wrangler override only; cost flag on activation)Must be on the vendor’s current model list at instantiation. [FACT-CHECK: model identifiers and availability are as of May 2026]
mcp_servers_allowedbroker-policy-library, broker-research-archive, broker-financial-newsMust be a subset of the broker manifest; the runtime refuses a deprecated server (MCP and connector broker)
built_in_toolsRead, Grep, WebFetchBash and Write usually disabled outside development zones
cost_ceiling_usd_per_day / _per_month50 per analyst / 1,500 aggregateA zero budget is invalid: no budget, no production envelope. Enforced at the model gateway
extended_thinking_budget_tokens50,000 per session0 disables extended thinking
data_classes_permittedPublic; Internal; Confidential non-customer; Confidential customer-financial via tokenisationMust be disjoint from the excluded list; an overlap is a registration error
data_classes_explicitly_excludedRestricted untokenised; Strictly confidentialRuntime denies, gateway denies, SOC alerts
network_egress_allow_listFive named financial-data vendor domains
gateway_route_refThe gateway route that resolves to workspace identity and the policy decision point

Memory

FieldValueNote
conversation_memoryper_sessioncross_session is not permitted at T2 and above
retrieval_scopeCurated research index; policy libraryEach entry resolves to a registry record with classification and retention
vector_store_accessEnabled, research index onlyEnabled with no named index is a registration error
prompt_cache_ttl1h for system prompt and research-style guideBook 3: one-hour cache writes cost 2× input, reads 0.1×; in the composite this is the difference between about USD 2,000 and USD 200 a month. Use 1h only where the write pays back within the hour. [FACT-CHECK: prompt-caching TTL options and multipliers against current Anthropic pricing]
persistent_storagenoneDeliberately no Files API, which Book 3 records as not eligible for zero data retention; tokenisation carries client context instead [FACT-CHECK: current ZDR eligibility of the Files API]
memory_eviction_rulesevict_on_session_endEvery Memory field must name when eviction fires
memory_audit_obligationtrueAlways true at T2 and above
cross_conversation_retentionfalseOverriding to true at T2+ needs second-line co-signature, a time-bound eviction trigger and a sensitive-disclosure clause

Monitoring

FieldValueNote
hook_bundle_refWorkstation hook bundle extended for SDK with SubagentStart / SubagentStop
evaluation_suite_ref200-sample synthesis-quality test set, threshold 92% match on golden answersMust resolve to a registered set with sample size and threshold
soc_detection_rulesToken spend > 200% of baseline; cross-class data retrieval; off-allow-list MCP attempt; subagent recursion depth > 3; configuration change from a non-policy sourceMinimum: at least one denial-rate rule and one cost rule
evidence_capture_refAudit Evidence Schema, nine questions on every material actionAudit evidence
review_cadence / review_owner_roleMonthly (T2) / the wranglerWeekly at T3+, quarterly at T0 and T1
incident_playbook_refHallucinated synthesis used as a citationIncident response
suspension_contact_role / _phoneWrangler out of hours, deputy head of research; hotline tested quarterly (last test logged)The hotline is tested at every quarterly review and the test is logged in the registry
model_upgrade_clauseRe-run evaluation within 24 hours; wrangler attests within 2 working days; Virtual CAB notified within 5 working days; Mandate re-validatedModel governance

What the envelope did in sixty days

Six near-misses, each stopped by a different clause: three Mandate refusals (autonomy ceiling draft), one Means denial (an off-allow-list MCP attempt), one Memory eviction (cross-conversation retention false prevented a disclosure carrying over between sessions), and one Monitoring suspension (a token-spend spike at 03:14 on a Saturday). The one minor incident was a poisoned external email summarised into context that triggered an off-list URL fetch; the Means denial fired and the chain was documented internally as below every reporting threshold. Book 3’s framing for the board: the agent that gets suspended six times in sixty days is the agent that is being run.

Validation and lifecycle checks

StageCheck (Book 3 App D §5)
RegistrationSchema (required and triggered conditional fields); cross-reference (data classes in the dictionary, MCP servers in the broker manifest, wrangler identity active, model on the vendor list); FRIA where triggered; risk-tier routing (T0/T1 default lane; T2 second-line co-signature; T3 plus the Agent Control Board; T4 prohibited without executive exception)
Every invocationMeans allow-list fails closed with a structured error to the SIEM; Memory eviction executes at its trigger and is logged; nine-question evidence emitted with envelope_id as a correlation key
Review (per review_cadence)Wrangler attests Mandate, minimal tools, memory and monitoring; business owner signs Mandate; DPO signs Memory; security architect signs Means; summary to the Virtual CAB
RetirementWrangler signs decommissioning; identity revoked; tool bindings and gateway routes removed; memory purged per schedule; registry status retired, record kept seven years; deletion is forbidden

Lifecycle operations are on agent lifecycle.

Part 3: two contrasting envelopes

A Claude Code review subagent (Bank A, Book 3 App D §4)

A subagent that checks incoming pull requests against the bank’s recipe (secure-development standards, regulatory-coding controls, data-handling patterns) and leaves a structured comment for a human reviewer. It runs inside Claude Code via the Agent SDK, in git worktree isolation. Envelope BANK-A-RISKTECH-CODEREVIEW-007. The differences from the research envelope are the lesson:

FieldResearch synthesis (T2)Code review subagent (T1)Why it differs
autonomy_ceilingdraftrecommend: posts a comment, never approvesEvery PR keeps a named senior reviewer
deployment_routedirect_api_apimfoundry (set through managed settings)Route follows the estate, not the use case [FACT-CHECK: Claude Code routing to Microsoft Foundry via managed settings, as of May 2026]
model_fallbackNamed, override onlynullNo fallback at T1
mcp_servers_allowedThree research brokersThe bank’s recipe broker; security-controls broker
built_in_toolsRead, Grep, WebFetchRead, Grep, Edit (worktree only)Plus worktree_isolation: true
Cost ceilingsUSD 50/day per analyst; 1,500/monthUSD 100/day per engineer; 25,000/month
Extended thinking50,000 tokens0 (disabled)
Data classes permittedUp to confidential customer-financial via tokenisationInternal only (bank source code); Public explicitly excludedThe subagent must not pull from public sources
EgressFive vendor domainsInternal Git and artefact repository only
prompt_cache_ttl1h5mEngineers iterate fast
EvictionSession endSession end or PR close
Evaluation200 samples, 92% match150 historical PRs with reviewer-validated comments, 90% precision, re-run nightlyHeld at 91–93% in the first 30 days
SOC rulesIncludes recursion depth > 3Includes worktree escape
Review cadenceMonthlyQuarterly
Model-upgrade clause24 h / 2 days / 5 days; Mandate re-validated48 h / 5 days / 10 days; Mandate not re-validated unless the tool surface changes

At 30 days (composite). About 180 PRs a week. The worktree-escape rule fired three times, each an engineer pointing the subagent at a working-tree path by mistake; denied and corrected. The off-list MCP rule fired once, on a deprecated recipe-broker name an engineer had cached locally; denied and purged. Coding-agent workstation controls are on coding agent workstation.

An HR policy assistant (Bank D, Book 1 Ch 2 §6.1)

Book 1’s envelope is the vendor-neutral one-page template, signed by four countersignatories (business owner with a CIO/CTO delegate; lead architect with the CISO; data owner with the DPO; operations lead with the head of audit). Its worked example is a well-bounded HR assistant at a composite specialist bank of about 3,500 staff.

SectionFilled values
MandateScope: colleagues’ HR policy questions (leave, benefits, pay enquiries, standard people processes). Outcome: grounded answer citing the HR policy document; route to a human HR specialist when the confidence threshold rejects. Exclusions: no legal advice, no interpretation of edge-case grievances, no transactions (cannot change a leave balance, approve a request or alter pay). Autonomy: advisory only. Override: nothing to reverse because nothing transacts; escalation to a named HR partner is one click
MeansOwn service-principal identity, distinct from the caller’s. Read-only HR policy knowledge base plus a few employee-scoped read APIs returning only the caller’s data (today’s leave balance, pay date). No write APIs. Credential lifetime: minutes per session. Excluded: payroll write, HRIS write, performance-management read
MemoryReads: policy index plus per-colleague reads at session time. Writes: audit log of question-answer pairs and ephemeral session context. Nothing colleague-identifiable persists beyond the audit window; anonymised aggregates feed the evaluation harness. No other agent reads this memory. Data ceiling: personal data, no special category
MonitoringTrace fields: question, retrieved citations, answer, confidence, escalation flag, latency, model version, prompt-template version. Evaluations: weekly factuality against a curated set; monthly drift; continuous policy-conformance check (does it cite a policy; is the policy current; does the answer contradict the text). Alerts: factuality regression > 2 percentage points to the wrangler; escalation rate > 1.5× baseline to HR operations. Wrangler: HR-tech operations lead; second line: HR director

Book 1’s point: the high first-contact resolution comes from how tight the envelope is, not how powerful the model is. Its readiness test (the Monday-morning test): when an incident occurs, can the on-call lead open the envelope in under two minutes and decide whether the incident is inside it (a defect to fix) or outside it (a containment and disclosure event)?

Book 1 and Book 3 envelopes are layers, not rivals

AspectBook 1 (vendor-neutral)Book 3 App D (Claude variant)This page follows
FormOne-page template, plain-English fields, four countersignatoriesMachine-readable registry record with enums and validation rules, read by runtime, gateway and SOCBook 1 as the governance document; Book 3 as its enforced form
Autonomy vocabularyAdvisory / controlled execution / autonomous standard-changerecommend / draft / execute_with_approval / execute_autonomouslyBook 3 for Claude agents; map both to Book 2’s agency classes on bounded agent
Review cadenceAnnual baseline; re-sign on material change; a model version bump within a family at the same evaluation results is not materialBy tier (weekly T3+, monthly T2, quarterly T0/T1); every model upgrade re-runs the evaluation on a clockBook 3 for production Claude agents. Book 1’s “not material” judgement predates Book 3’s model-upgrade clause and should not be used to skip it

Takeaways

  • Divide the request by control flow and responsibility before choosing a model or framework; the product-shape guide gives the sequence, and workflow and bounded agent describe the shapes.
  • Write the design record before the first branch, with rejected classes and their reasons.
  • Register one envelope per running agent, with a budget, an evaluation threshold, alert thresholds and a tested suspension contact, before production.
  • Compare envelopes across agents: the differences (ceiling, egress, data classes, cadence) should each trace to a reason in the design record.

Related worked examples: a migration through the factory, and an inspection answered from the record, which shows how the envelope and evidence fields are read back by a supervisor.

Source material

  • Book 2 (The AI Delivery Factory), Chapter 22, The Shape of an AI Product (§5.4 classification line; §5.5 design record; §6 the research crew that became an assistant; §8 risks)
  • Book 3 (Deploying Claude Safely), Appendix D, Bounded-Agency Envelope Template (Claude Variant) (§2 schema; §3 API agent record; §4 Claude Code subagent record; §5 validation and lifecycle)
  • Book 1 (Building Organizations for the Age of AI), Chapter 2, The Bounded Agency Thesis (§5 envelope template; §6.1 HR assistant at Bank D; §7 Monday-morning test)
  • Book 2 (The AI Delivery Factory), Appendix C, The Pattern Catalogue (pattern names referenced by design-record field 4)
  • Research Paper 1, Taxonomy for a Reuse-Driven AI Factory (§7.5 the design record)

← All AI Delivery Factory guidance

On this page

  1. Part 1: the product-shape decision
  2. The request
  3. Apply the decision sequence
  4. Preserve the human boundary
  5. What was rejected, and the revisit trigger
  6. Outcome (composite, read the following spring)
  7. Part 2: writing it down
  8. The classification line
  9. The design record (ten fields)
  10. The envelope record: the Claude variant
    1. Mandate
    2. Means
    3. Memory
    4. Monitoring
    5. What the envelope did in sixty days
    6. Validation and lifecycle checks
  11. Part 3: two contrasting envelopes
  12. A Claude Code review subagent (Bank A, Book 3 App D §4)
  13. An HR policy assistant (Bank D, Book 1 Ch 2 §6.1)
  14. Book 1 and Book 3 envelopes are layers, not rivals
  15. Takeaways

All pages in Worked examples →