Shipping a Feature End-to-End with Claude Agents

Most writing about agentic development stays abstract — orchestrators, subagents, context windows. This post does the opposite. We are going to walk one realistic feature from a fuzzy stakeholder request all the way to shipped, verified code, using Claude Code and the agentic patterns that hold up in practice. The goal is to show the actual texture of the work: where the human leads, where the agent runs, and where things go sideways.

The scenario is deliberately ordinary, because ordinary is where most engineering time goes. A B2B SaaS team needs to add a "saved filter views" feature: users want to save a combination of table filters, name it, and reload it later. It touches the database, the API, and the frontend. A traditional implementation is a few days of focused work. Let us see how it goes when an agent does the bulk of the typing and a human does the thinking.

Step 1 — Turning a vague request into an executable spec

The request arrives as one sentence: "Let people save their filters." That is not something you hand to an agent yet — it is something you hand to an agent and get back three plausible-but-wrong implementations. The first and most important human step is specification. You decide the actual behavior: a saved view stores the active filter set plus a sort order; views are per-user and private; names must be unique per user; deleting a view is a soft delete; loading a view replaces the current filters.

You write this down — in the issue, and as guidance Claude will read. You also decide, up front, how you will know it works: the API returns the right shape, the unique-name constraint is enforced, and a user can round-trip save-then-load. Defining "done" before the agent starts is what separates a clean run from an afternoon of re-prompting. The spec is short, but it removes the ambiguity that causes agents to guess wrong.

Step 2 — Letting Claude Code plan across the codebase

flowchart TD
  A["Vague request"] --> B["Human writes spec & success criteria"]
  B --> C["Claude Code plans across DB, API, UI"]
  C --> D["Subagents implement migration, endpoint, frontend"]
  D --> E["Run tests & evals"]
  E -->|Fail| F["Agent iterates on failures"]
  F --> E
  E -->|Pass| G["Human reviews diff & verifies"]
  G --> H["Ship to staging then prod"]

With the spec in hand, you point Claude Code at the repository and ask it to plan, not yet to code. It reads the existing schema, the API conventions, and the frontend table component, then proposes a plan: a new saved_views table with a migration, a set of CRUD endpoints following the existing controller pattern, and a UI affordance in the filter bar. This planning step is where reading the codebase pays off — the agent reuses your conventions instead of inventing new ones, because it has the actual files in context.

You review the plan and correct one thing: the agent proposed a hard delete, but your spec said soft delete. This is the moment that justifies the whole spec exercise — the deviation is caught in seconds, before any code exists. You approve the corrected plan. For a feature this size you let Claude Code fan out: parallel subagents take the migration, the endpoint, and the frontend, each with the plan and conventions as shared context. Multi-agent work costs more tokens, so you use it because the subtasks are genuinely independent.

Step 3 — Implementation, tests, and the iteration loop

The subagents produce a migration, an API layer, and a frontend control. None of it is trusted yet. The agent runs the test suite and hits two failures: the unique-name constraint throws an ungraceful 500 instead of a clean 409, and the frontend sends filters in a slightly different shape than the endpoint expects. This is the agentic loop working as designed — the agent reads the failing test output, diagnoses the mismatch, and fixes both, then re-runs until green.

You watch this loop rather than driving it, but you do not switch off. When the tests pass, you look at why they pass. One test asserts the endpoint returns 200 on save — but does it assert the saved record actually contains the sort order? It does not. You add that assertion yourself (or ask the agent to, then verify it). This is the verification discipline that catches confident wrongness: a green suite that checks the wrong thing is worse than a red one, because it lulls you.

Step 4 — Review, verification, and shipping

Now the human-led review. You read the full diff, not as a formality but hunting for the specific failure modes agents produce: a plausible-but-subtly-wrong query, a missing authorization check (does the load endpoint confirm the view belongs to the requesting user?), an N+1 introduced by convenience. Here you find one real issue — the list endpoint does not filter by user, so any authenticated user could load anyone's views. The tests passed because no test covered cross-user access. You add that test, the agent fixes the query, and now the eval suite is stronger than before.

With the diff clean and the evals — including the new authorization and round-trip checks — passing, you ship to staging. You manually round-trip the feature: save a view, reload the page, load the view, confirm the filters and sort restore exactly. It works. You promote to production behind a flag, watch logs and metrics for errors, and ramp. The whole feature took an afternoon instead of three days, but the time saved was in typing — the judgment, specification, and verification were entirely human, and that is exactly why it shipped safely.

What this walkthrough teaches about agentic development

The pattern generalizes. The human owns the bookends — the spec and success criteria at the front, the review and verification at the back — while the agent owns the middle: planning across the codebase, implementing, and iterating against tests. The failures the agent introduced (hard delete, missing auth filter, weak assertions) were all caught because a human defined "done" precisely and refused to trust a green checkmark. Remove either bookend and the speed becomes a liability.

Notice also what made the run smooth: a codebase Claude could read, conventions it could follow, a test suite it could iterate against, and a spec that removed ambiguity. Teams that invest in those four things get reliable agentic runs; teams that point an agent at a vague request and an untested codebase get fast, confident, wrong code.

Frequently asked questions

How much of an agentic feature should a human write?

Often very little code, but all of the specification, success criteria, and final verification. The agent's leverage is in planning and typing; the human's leverage is in defining correct behavior and proving the result actually meets it. Skipping the human bookends is where agentic speed turns into shipped bugs.

When should I use multiple Claude Code subagents on one feature?

When a feature splits into genuinely independent subtasks — like a migration, an endpoint, and a frontend control that can be built in parallel. Multi-agent runs use several times more tokens than a single agent, so reserve them for real parallelism rather than reaching for them by default.

Why did the tests pass even though there was a security bug?

Because no test covered cross-user access — a green suite only proves the behaviors it actually asserts. This is the central agentic verification lesson: review the diff for the failure modes tests miss, then turn every gap you find into a new eval so it cannot recur.

What makes an agentic run go smoothly versus painfully?

Four things: a readable codebase, consistent conventions the agent can follow, a test suite it can iterate against, and a precise spec that removes ambiguity. Invest in those and runs are reliable; point an agent at a vague request and an untested repo and you get fast, wrong output.

From shipped features to answered calls

CallSphere applies this same spec-plan-verify loop to voice and chat: agents that take a real request, use tools mid-conversation, and book the outcome — with the verification discipline that keeps automated work trustworthy. See it in action at callsphere.ai.

Source & attribution: This is an independent, original explainer inspired by Anthropic's coverage on the Claude blog. Claude, Claude Code, Claude Cowork, Claude Opus, and the Model Context Protocol are products and trademarks of Anthropic. CallSphere is not affiliated with or endorsed by Anthropic.