The design unit is the operational decision, not the model call. The delivery unit is the smallest end-to-end solution slice that can prove a customer or internal outcome with representative data, real controls, and a usable work surface.
1. Trace requirements to the solution
Create one row for every functional requirement:
| Requirement | Domain objects/state | Context and logic | Action/effect | Security | User surface | Verifier | Telemetry | Owner |
|---|---|---|---|---|---|---|---|---|
| Named decision | Entities, relationships, lifecycle | Sources, freshness, rules, models | Tool, precondition, postcondition | Actor, scope, approval | Artifact, review, interruption | Outcome or readback | State, policy, effect, outcome | Accountable role |
If a component cannot be traced to a requirement, justify or remove it. If a requirement cannot be traced to state, action, verification, and ownership, the design is incomplete. Palantir's solution-design method uses object models, lifecycle diagrams, enrichments, and interface expectations to preserve this traceability. R26-41
When the approved workflow falls into a recurring delivery shape, use the operational solution portfolio to start from a bounded vertical slice:
- Select one business-flow pattern by its trigger, decision, action, and accepted outcome.
- Apply an industry profile only where it adds concrete domain objects, evidence, authority, risk, or operating measures.
- Select one primary horizontal accelerator by the dominant technical failure boundary.
Copy the relevant canonical starter packet into the delivery repository and replace every assumption with target-specific evidence. Flow patterns, vertical profiles, and horizontal accelerators are design inputs—not release artifacts or substitutes for field observation.
Controls: FDE-001, ARC-001, DEL-001.
2. Model data, logic, action, and security together
Use the operational-ontology template to represent:
- Data: real business entities, identity, relationships, source authority, time, classification, and evidence lineage
- Logic: deterministic rules, calculations, policies, optimizers, models, and agent judgment
- Action: lifecycle transitions, tool contracts, effects, approvals, compensation, and readback
- Security: actor identity, caller authority, tenant, purpose, data scope, policy, and audit
This is the portable architectural lesson from Palantir's decision-centric Ontology: model the operational world and the changes allowed within it, not merely the documents retrieved by the model. It does not require a specific vendor or graph technology. R26-42
Control: ARC-002.
Promote the discovery assessment into a validated data-context manifest. Its four planes, decision-critical profile, preparation versions, labels, output records, economics, and operating response must match the selected mechanisms. Use the data preparation and context pipeline when parsing, reconciliation, redaction, chunking, indexing, features, or derived views are material production components. CTX-006 through CTX-009.
3. Select the smallest sufficient decision mechanism
Evaluate these in order when they fit the decision:
- Deterministic software
- Optimization or constraint solver
- Classical ML or statistical model
- Search and retrieval
- One model call with typed input and output
- A coded workflow with model-assisted steps
- One bounded agent with tools and stop conditions
- Multiple workers only across a genuine context, permission, ownership, specialization, or latency boundary
- Human review where the decision remains weakly verifiable or high stakes
Record the baseline, the measured failure that justifies each increase in agency, full cost, fallback, and monitor in the intelligence-selection record. OpenAI and Anthropic both recommend incremental complexity and a single-agent baseline. R26-40 R26-56
Controls: ARC-004, ARC-005.
4. Use a layered production architecture
The enterprise agent platform blueprint separates:
- Professional work surface and persistent artifact
- Durable workflow and agent harness
- Operational domain and context assembly
- Typed tools and effect gateway
- Identity, authorization, secrets, egress, and sandbox controls
- Evaluation, observability, and evidence stores
- Delivery control plane for versions, branches, promotion, and rollback
The model proposes. Trusted software validates, authorizes, executes, persists, and verifies.
The product layer bridges intelligence and real work. Design it as one coherent chain:
operator trigger and work surface
-> governed context and decision mechanism
-> typed proposal or action boundary
-> persistent evidence and review state
-> source-of-truth result, outcome signal, and operating feedback
Do not reduce the product to output tokens. The application needs workflow-native context capture, task-shaped tools, a durable review artifact, and feedback paths that users can understand and operate. The internal applied-AI team can make those elements reusable; the operating team remains responsible for the local decision and policy. ADP-001, ARC-002, FDE-004.
5. Build a vertical slice
The first slice must cross the real boundaries of the workflow:
representative trigger
-> authorized context
-> bounded reasoning
-> typed proposal
-> policy decision
-> staged or simulated effect
-> source-of-truth verification
-> persistent user artifact
-> trace and accepted-outcome event
Do not begin with every integration, user segment, or exception. Use one representative segment with the final identity, tool, telemetry, evaluation, and review patterns. Replace only those production dependencies that cannot safely be exercised, and record every substitute.
Controls: ARC-001, TOL-001, REL-003, OPS-001.
Map complexity only when it changes delivery quality
When cross-team dependencies repeatedly slow discovery, incident response, onboarding, or release review, create separate derived software and operational views. Use the evidence graph and change-intelligence blueprint, a system-map manifest, and a change-impact assessment. These views route people to likely impacts; they never replace source-of-truth policy, tool authorization, evaluation, approval, or readback. CTX-001, CTX-002, CTX-004, OPS-007.
6. Treat the solution as one release unit
Version these together:
| Layer | Required identity |
|---|---|
| Data and context | Data-context manifest, source/schema revisions, quality snapshot, preparation/index/feature versions, label authority, and output-record contract |
| Domain | Ontology/domain-model version and migration |
| Behavior | Behavior-bundle ID/version/digest that binds the model, prompt/instruction, route, harness, context policy, complete tool membership, and guardrail |
| Capabilities | Tool contract plus exact capability-manifest ID/version/digest, implementation provenance, authorization, network policy, and lifecycle |
| State | Workflow schema, checkpoint, migration, retention |
| Assurance | Threat model, evaluation suite, grader, world fixture |
| Experience | User surface, evidence packet, approval contract |
| Operations | Runtime, deployment, alert, runbook, rollback |
Isolate cross-resource changes, review compatibility, and test them end to end. Merge is a source-control event; deployment is an independently gated production event. R26-44
Every tool addition or change requires a new capability decision, affected-route evaluation, and an update to behavior-bundle membership and the solution-release digest when the current model/agent release contract applies. A semantic version without exact artifact and authority digests is not a release binding.
The current behavior-bundle, evaluation-report, and solution-release contracts apply when model or agent behavior is selected. A deterministic, optimization, or classical-ML-only route retains equivalent mechanism, test, provenance, deployment, rollback, and operating evidence without placeholder agent components.
Controls: DEL-001, DEL-002, OPS-007.
7. Deliver with users, not toward them
Use a short evidence loop:
observe case -> update requirement -> build slice -> replay cases
-> operator reviews artifact -> measure friction and outcome -> decide next slice
Recommended cadence:
| Cadence | Participants | Decision |
|---|---|---|
| Daily field sync | Delivery lead, technical owner, operator representative | New facts, blockers, risky assumptions |
| Twice-weekly case review | Operators, domain expert, delivery lead | Correctness, exceptions, UX, evidence quality |
| Weekly outcome review | Operational owner, product, technical, risk | Scope, metric movement, stop/continue decision |
| Release gate | Technical, operational, risk, service owners | Evidence-based promotion or rollback |
Keep a product backlog and an assurance backlog. A feature is not complete when its behavior, security, observability, support, or training remains undefined.
Separate the customer-specific configuration backlog from the reusable product/platform backlog. A recurring field signal becomes a candidate shared improvement only after it is sanitized, its recurrence is evidenced without moving customer data, and its destination passes the normal design, evaluation, release, and rollback gates. Use the field-learning register; the FDE or delivery team does not unilaterally convert a local workaround into a platform feature. FDE-004, DEL-002.
8. Predeclare adoption evidence
Before pilot entry, define the adoption metric independently of the observed result:
| Field | Required definition |
|---|---|
| Eligible denominator | Users or workflow opportunities that could validly use the agent, with exclusions and deduplication |
| Baseline | Value, status, as-of date, and comparison population |
| Target | Dated threshold for the named segment |
| Guardrail | Maximum abandonment, rework, unsafe approval, reviewer load, or support burden |
| Window | Event time, timezone, late-event rule, and reporting period |
| Source | Authoritative event/query, schema, and revision |
| Owner | Role accountable for definition, data quality, review, and rebaseline |
Instrument eligible exposure, completion, accepted outcome, override, rejection, abandonment, reviewer wait, and support events during the pilot. A denominator, source, window, or exclusion change creates a new metric revision; it is not an in-place correction to an unfavorable result.
Diagnose the first broken transition before prescribing training or expanding the cohort:
| Transition | If it is weak | First response |
|---|---|---|
| Eligible -> exposed | Access, routing, integration, data readiness, or eligibility may be wrong | Repair the path or narrow the declared segment |
| Exposed -> completed | The surface, timing, instructions, trust, or safe fallback may be wrong | Observe cases and remove product friction |
| Completed -> independently accepted | Evidence, behavior, policy, or authority may be wrong | Repair context, mechanism, evaluation, or boundary |
| Accepted -> business effect | The workflow or baseline may be wrong | Revisit the product and value thesis |
| Business effect -> sustained value | Review, support, recovery, or operating cost may be too high | Constrain, redesign the operating model, or stop |
Record reason codes from actual cases rather than inferring motives from low usage. Run access and surface rehearsal before shadow; let operators lead the bounded cohort before canary; require the receiving team to reproduce the funnel and support path before handoff.
At pilot entry, re-confirm the predeclared graduation contract without changing thresholds after seeing results. Technical performance, operator acceptance, adoption, business value, full economics, and production readiness remain separate decisions. A strong result in one dimension cannot compensate for a failed authority, safety, ownership, or value gate.
Use the delivery and adoption plan.
Controls: FDE-003, VAL-001, VAL-002, OPS-006.
9. Build the professional work surface
An agent interaction must resolve into a durable artifact: case file, spreadsheet, document, notebook, review table, ticket, code change, or operational application. The user needs local evidence, editable state, alternatives, uncertainty, and permitted actions. Chat may initiate or explain work; it should not be the only review and accountability surface. R26-43 R26-54
Control: ADP-001.
10. Manage prototype debt explicitly
Before shadow use, classify every shortcut:
| Shortcut | Required disposition |
|---|---|
| Mock or copied data | Replace with governed source or keep the slice non-production |
| Broad user/service credential | Replace with actor-aware least privilege |
| Prompt-only rule | Move enforcement below the model |
| Live mutable eval | Replace with versioned replay environment |
| Manual hidden step | Model as state, tool, review, or named operator procedure |
| Best-effort retry | Add timeout, idempotency, terminal state, and recovery |
| Unstructured approval | Bind reviewer, proposal digest, policy, expiry, and effect |
| Ad hoc deployment | Add compatible release manifest, canary, rollback, and owner |
11. Start adoption and harness handoff at pilot entry
At pilot entry, bind every solution artifact to its authoritative URI, version or digest, upstream lineage, change owner, receiving owner, promotion path, and retention policy. The accountable technical and service owners pair on harness and behavior changes, adoption queries, evaluation construction, release, alerts, incidents, and support. Before bounded production, they lead those paths with the delivery lead observing. If the building team remains the service owner, a designated backup exercises them without relying on the original builder. The final handoff is proof of exercised capability, not a document dump.
Use the delivery and adoption plan and customer enablement handoff.
Control: ADP-002.
Delivery exit gate
- [ ] Every requirement traces to domain, logic, effect, security, UX, verifier, telemetry, and owner.
- [ ] The deterministic/single-call/workflow baseline is recorded.
- [ ] The vertical slice exercises representative data and final control boundaries.
- [ ] Tools, policies, state, effects, and failure behavior are machine-readable and tested.
- [ ] Operators can inspect, correct, pause, reject, escalate, and resume work.
- [ ] Cross-resource versions form one compatible release candidate.
- [ ] Replay, shadow, safety, cost, and adoption evidence meet their separate thresholds.
- [ ] Adoption denominator, baseline, target, guardrail, window, source, and owner were fixed before pilot evidence was reviewed.
- [ ] Funnel breaks were diagnosed from representative cases and owned actions, not explained away as generic resistance.
- [ ] Reviewer and support capacity were exercised for the proposed cohort and have declared limits.
- [ ] Pilot graduation gates, evidence cutoff, decision date, and stop or redesign path remained frozen or changed through an explicit revision.
- [ ] Prototype debt has an owner and blocking disposition.
- [ ] Customer owners have exercised harness change, measurement, release, support, and incident paths.
- [ ] Production artifacts resolve to owned immutable versions or digests with upstream lineage and promotion paths.
Delivery anti-patterns
- Architecture built around the model API instead of the operational decision
- Retrieval index presented as a full domain and action model
- Multi-agent topology before a serial baseline
- Code tested independently from data, policy, tools, UX, and evaluators
- Chat-only professional review
- Pilot velocity used as proof of maintainability or value
- Adoption measured from registered users or runs without a declared eligible denominator
- Unfavorable adoption hidden by changing the window, exclusions, or event source in place
- Late security, operations, adoption, or handoff workstream