Devin SDK example: Anthropic agent patterns

Source: anthropic-patterns

The patterns from Building effective agents with Devin as the agent:

Pattern	Run	What it shows
Prompt chaining	`bun run chaining`	Sequential steps in one session, a typed structured-output gate between them
Routing	`bun run routing`	A classifier turn whose zod-typed result dispatches to specialized handlers
Parallelization	`bun run parallelization`	Independent sessions under `Promise.all` — sectioning and majority voting
Orchestrator-workers	`bun run orchestrator-workers`	An orchestrator plans subtasks, workers run in parallel, the orchestrator merges
Evaluator-optimizer	`bun run evaluator-optimizer`	A generator/evaluator loop driven by a typed pass/feedback verdict

The glue in every pattern is structured output: session.run(prompt, { output: zodSchema }) instructs the agent to reply with matching JSON, validates the reply, and gives you a fully typed outcome.output to branch on in plain TypeScript. Nothing here is Devin-specific beyond that one call, which is the point: control flow stays in your code, the model fills in one step.

evaluator-optimizer.ts (excerpt)

const Verdict = z.object({ pass: z.boolean(), feedback: z.string() });

for (let round = 1; round <= 3; round++) {
  const verdict = await evaluator.run(
    `Grade this against the criteria. No tools.\nCriteria: ${criteria}\nDraft:\n${draft}`,
    { output: Verdict },
  );
  if (verdict.output.pass) break;
  // The generator session keeps its own history; just feed it the feedback.
  draft = (
    await generator.run(
      `Revise per this feedback, code only, no tools:\n${verdict.output.feedback}`,
    )
  ).text.trim();
}

What you should see

Model text varies run to run, but the shape is stable. bun run evaluator-optimizer prints the grading loop converging:

round 1: pass=false — The implementation … does not swap lo and hi when lo > hi …
round 2: pass=true — It is a single line and correctly handles lo > hi …

final draft:
const clamp = (n, lo, hi) => Math.max(Math.min(n, Math.max(lo, hi)), Math.min(lo, hi));

and bun run parallelization prints three independent one-sentence reviews followed by votes safe=[false, false, false] -> majority: unsafe.

Run

git clone https://github.com/COG-GTM/devin-sdk-examples
cd devin-sdk-examples/anthropic-patterns
bun install                    # @cognition-ai/sdk from npm (the pinned beta)
export DEVIN_API_KEY=cog_…
bun run chaining

Every pattern opens with a bare createDevin(), so it runs in the cloud. Pass createDevin({ cwd: process.cwd() }) to run through the bundled CLI instead (the patterns are chat-only, so nothing is written there), or createSession({ repos: ["your-org/your-repo"] }) to give a cloud session a repository to work in.

Launch and structured output

How run(prompt, {output}) validates the reply and what happens when it doesn’t match.

Ticket enricher

One of these patterns deployed as a webhook, with a queue and a typed result.

Get started

Sessions

Integrations

Examples

SDK reference

API client

Resources

Devin SDK example: Anthropic agent patterns

What you should see

Run

Launch and structured output

Ticket enricher

​What you should see

​Run

Launch and structured output

Ticket enricher

What you should see

Run