diff --git a/src/automation/durableRunCoordinator.retryBackoff.test.ts b/src/automation/durableRunCoordinator.retryBackoff.test.ts
new file mode 100644
index 00000000..4def70b0
--- /dev/null
+++ b/src/automation/durableRunCoordinator.retryBackoff.test.ts
@@ -0,0 +1,95 @@
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import type { PipelineResult } from '../agents/pairPipeline.js';
+import type { TaskItem } from '../orchestration/decisionEngine.js';
+import { DurableRunCoordinator, retryAtFor } from './durableRunCoordinator.js';
+
+const roots: string[] = [];
+
+function dbPath(): string {
+  const root = mkdtempSync(join(tmpdir(), 'openswarm-coordinator-backoff-'));
+  roots.push(root);
+  return join(root, 'automation.db');
+}
+
+function task(id: string): TaskItem {
+  return {
+    id,
+    issueId: id,
+    issueIdentifier: id,
+    source: 'linear',
+    title: `Task ${id}`,
+    priority: 2,
+    createdAt: Date.now(),
+    linearState: 'Todo',
+    linearProject: { id: 'project', name: 'Repo' },
+  };
+}
+
+function failure(finalStatus: PipelineResult['finalStatus']): PipelineResult {
+  return {
+    success: false,
+    sessionId: 'session-1',
+    stages: [],
+    finalStatus,
+    totalDuration: 100,
+    iterations: 1,
+  };
+}
+
+afterEach(() => {
+  vi.useRealTimers();
+  for (const root of roots.splice(0)) rmSync(root, { recursive: true, force: true });
+});
+
+// A task that fails every time was re-claimed as soon as its attempt ended and kept
+// a slot for good: 11 issues took 74% of 78 attempts in four hours while 80 others
+// waited (AGT-4673). The wait now doubles from the fourth attempt.
+describe('retry backoff for a task that keeps failing (AGT-4673)', () => {
+  it('keeps the 30-minute retry for three attempts, then doubles it up to six hours', () => {
+    const failed = failure('failed');
+    const minutes = (attemptNo: number) => retryAtFor(failed, 0, attemptNo) / 60_000;
+
+    expect([1, 2, 3].map(minutes)).toEqual([30, 30, 30]);
+    expect([4, 5, 6].map(minutes)).toEqual([60, 120, 240]);
+    expect([7, 8, 40].map(minutes)).toEqual([360, 360, 360]);
+    expect(retryAtFor(failed, 5_000)).toBe(5_000 + 30 * 60_000);
+    expect(retryAtFor(failure('rejected'), 0, 5) / 60_000).toBe(120);
+  });
+
+  it('ramps an infrastructure error from fifteen minutes to a two-hour cap', () => {
+    const infra = failure('infra_error');
+    const minutes = (attemptNo: number) => retryAtFor(infra, 0, attemptNo) / 60_000;
+
+    expect([1, 3].map(minutes)).toEqual([15, 15]);
+    expect([4, 5, 6, 9].map(minutes)).toEqual([30, 60, 120, 120]);
+  });
+
+  it('does not slow a rate limit by its attempt count', () => {
+    const limited = failure('rate_limited');
+    expect(retryAtFor(limited, 1_000, 12)).toBe(1_000 + 60_000);
+    expect(retryAtFor({ ...limited, rateLimitResetsAt: 9_000 }, 1_000, 12)).toBe(9_000);
+  });
+
+  // retryAtFor is only half of it: the coordinator has to pass the run's own attempt
+  // count, or every failure keeps the first-attempt delay.
+  it("applies the ramp from the run's own attempt count when a failure is recorded", async () => {
+    vi.useFakeTimers();
+    vi.setSystemTime(1_000);
+    const coordinator = new DurableRunCoordinator({
+      mode: 'primary', dbPath: dbPath(), instanceId: 'ramp-owner',
+    });
+    const waits: number[] = [];
+    for (let attempt = 1; attempt <= 5; attempt++) {
+      const startedAt = Date.now();
+      await coordinator.execute(task('AGT-RAMP'), '/repo', async () => failure('failed'));
+      const run = coordinator.getRun('AGT-RAMP');
+      waits.push(((run?.retryAt ?? 0) - startedAt) / 60_000);
+      vi.setSystemTime((run?.retryAt ?? 0) + 1);
+    }
+    expect(waits).toEqual([30, 30, 30, 60, 120]);
+    coordinator.close();
+  });
+});
diff --git a/src/automation/durableRunCoordinator.ts b/src/automation/durableRunCoordinator.ts
index 49d0ace3..3d7da656 100644
--- a/src/automation/durableRunCoordinator.ts
+++ b/src/automation/durableRunCoordinator.ts
@@ -155,8 +155,13 @@ export function retryAtFor(result: PipelineResult, now: number, attemptNo = 1):
     const exponent = Math.max(0, Math.min(16, attemptNo - 1));
     return now + Math.min(6 * 60 * 60_000, 5 * 60_000 * (2 ** exponent));
   }
-  if (result.finalStatus === 'infra_error') return now + 15 * 60_000;
-  return now + 30 * 60_000;
+  // A task that fails every time was re-claimed as soon as its attempt ended and
+  // kept its place at the top of the queue: 11 issues took 74% of 78 attempts in
+  // four hours while 80 others waited (AGT-4673). The delay stays at today's value
+  // for the first three attempts, then doubles with each one.
+  const ramp = 2 ** Math.max(0, Math.min(16, attemptNo - 3));
+  if (result.finalStatus === 'infra_error') return now + Math.min(2 * 60 * 60_000, 15 * 60_000 * ramp);
+  return now + Math.min(6 * 60 * 60_000, 30 * 60_000 * ramp);
 }
 
 function processIsAlive(pid: number): boolean {
@@ -933,7 +938,7 @@ export class DurableRunCoordinator {
       default: target = 'RETRY_AT'; break;
     }
     const transitioned = this.ledger.transition(claim, target, {
-      retryAt: target === 'RETRY_AT' ? retryAtFor(result, now) : null,
+      retryAt: target === 'RETRY_AT' ? retryAtFor(result, now, claim.attemptNo) : null,
       errorCode: result.finalStatus,
       errorMessage: detail,
       eventData: { sessionId: result.sessionId, finalStatus: result.finalStatus },