04 · Blueprint Sprint  ·  Reference Architecture Add-On

Design the agent.
Then ship it on a runtime that won't fight you.

A focused design engagement that maps your business processes, produces an implementation-ready agent experience — and hands engineering a Claude Managed Agents reference architecture they can deploy in days, not quarters.

Format
3-week design sprint
Best for
Teams ready to build their first or next agent
Output
Design + deployment spec
Runtime
Claude Managed Agents

The infrastructure problem just got solved. The design problem didn't.

Before

Months of plumbing

Building a production agent meant servers, sandboxing, state management, checkpointing, credential scoping, and an orchestration loop you'd own forever. Engineering owned the runtime; design owned a slide deck.

After

Anthropic owns the loop

Claude Managed Agents handles the runtime — sandboxed execution, persistent sessions, scoped tools, end-to-end tracing — billed by the millisecond at $0.08 per session-hour. The bottleneck moves upstream, to whether the agent is worth trusting in the first place.

Anthropic solved the runtime. We design the relationship that runs on top of it.

Five stages from process reality to deployment-ready spec.

Blueprint Sprint translates how your business actually works today into a coherent agent experience design — and hands engineering a Managed Agents reference architecture they can stand up against the documented contract.

Process Map

Walk current workflows, decision points, handoffs. Identify where agents add leverage and where they shouldn't go near.

Intent & Autonomy

What the agent owns, what stays human, what requires approval. Mapped against the AUX trust ladder.

Interaction Blueprint

Intent handshakes, confidence cues, escape hatches, memory in motion — the six AUX patterns wired into your specific flows.

Reference Architecture

Agent definition, environment spec, tool inventory, approval policy, session-runtime profile. Aligned to the Managed Agents contract.

Cost & Trust Forecast

Modeled session-hours, token volume, web-search load, cache strategy. Trust risks flagged before the first deploy.

Six artifacts your team can build from on Monday.

D1

Process Flow Maps

Current state and agent-augmented state, side by side. Decision points, handoffs, and friction made explicit.

D2

Agent Interaction Blueprint

Conversation flows, approval gates, escalation paths, fallback logic. Trust patterns named and placed.

D3

Intent & Autonomy Map

What the agent decides alone, what it proposes, what it must escalate. Tied to the AUX trust ladder.

D4

Managed Agents Reference Spec

Agent configuration, environment requirements, tool definitions, scoped permissions, session shape, event-stream expectations.

D5

Cost & Runtime Forecast

Modeled monthly cost across tokens, session-hours, and tools. Caching strategy. Sensitivity analysis on usage growth.

D6

Trust Risk Register

Where the design could break trust under real conditions, with mitigation paths and validation criteria for the build phase.

Four building blocks. Designed, not improvised.

Managed Agents runs on four primitives: Agent (the contract), Environment (the workspace), Session (the running conversation), Events (the message stream). Blueprint Sprint produces a design specification for each — so engineering implements decisions, not assumptions.

Cost Forecast Model

Three-axis billing, three design levers.

Managed Agents bills on tokens, session-runtime, and tool calls — simultaneously, within a single session. Each is a design choice before it's a finance line item.

monthly_cost = tokens(input + output, net of cache)
           + session_hours × $0.08 // metered to the millisecond, running only
           + web_searches × $0.01 // $10 per 1,000
Token strategy prompt caching · context discipline
Runtime strategy approval gating · session scoping
Tool strategy search budgeting · tool selection

Pricing reflects Anthropic's published Managed Agents beta rates as of Q2 2026. Forecasts include sensitivity bands for usage growth and cache-hit ratios.

Trust is a permission contract, not a feature.

Managed Agents lets you mix automatic execution with explicit human approval per tool, per action, per surface. We design that contract before code gets written — and frame it in language risk-averse stakeholders will actually sign off on.

Surface Action class Mode Trust pattern
Inbound research Read files, search the web, summarize Auto-run Confidence cues
Internal drafts Generate proposals, briefs, replies Auto-run Generative momentum
External comms Send email, post messages, publish Approval Clarify before commit
Systems of record Write to CRM, update ledgers, change state Approval Escape hatch + audit trail
Irreversible actions Payments, contracts, deletions Approval Clarify before commit · pushback

When this engagement is the right move.

Strong fit
  • You're committing to ship an agent in production within the next quarter
  • The workflow involves multiple decision points, handoffs, or systems of record
  • Stakeholders need confidence the agent won't break trust before approving the build
  • Your engineering team is capable but doesn't want to design the experience from scratch
  • You want to standardize how agents are designed before you have ten of them
Not the right fit
  • You're exploring whether to do anything at all — start with the Executive Seminar
  • The agent already exists and is in market — start with the Agent Experience Audit
  • You want a working agent demoed by end of week — Blueprint Sprint is a 3-week design engagement, not a build sprint (though we can run the build and operate it after, as a follow-on)

Bring us a workflow.
Leave with an agent worth deploying.

A 30-minute scoping call to map the workflow, identify the agent's surface area, and confirm Blueprint Sprint is the right engagement for where you are right now.

Start a conversation

Tell us about the workflow you want to turn into an agent. We'll scope it and come back with a view on what the sprint looks like for your context.

View All Engagements