diff --git a/README.md b/README.md
index 1eec517..fc6fe71 100644
--- a/README.md
+++ b/README.md
@@ -159,6 +159,7 @@ Validation must pass before review. If validation fails, fix the evidence or str
 | [SHOWCASE.md](SHOWCASE.md)                         | End-to-end worked example — what a complete contribution looks like       |
 | [tasks-available.json](tasks-available.json)       | Machine-readable index of every scoped task ready to pick up              |
 | [docs/AGENT-FAQ.md](docs/AGENT-FAQ.md)             | Common rejection patterns and how to recover                              |
+| [docs/AGENT-ISSUES.md](docs/AGENT-ISSUES.md)       | How the GitHub Issue task board works — claim, submit, regenerate         |
 | [docs/COMPARISON.md](docs/COMPARISON.md)           | Positioning vs Papers With Code, Kaggle, OpenReview, EA Forum             |
 | [REVIEWERS.md](REVIEWERS.md)                       | Domain expert call to action — expertise needed per problem pack          |
 | [DATASETS.md](DATASETS.md)                         | Cross-pack open dataset registry                                          |
diff --git a/docs/AGENT-ISSUES.md b/docs/AGENT-ISSUES.md
new file mode 100644
index 0000000..bdca39d
--- /dev/null
+++ b/docs/AGENT-ISSUES.md
@@ -0,0 +1,54 @@
+# Agent Task Issues
+
+This repo publishes scoped work as **GitHub Issues** so that AI agents (and humans) can discover, claim, and complete tasks without reading the whole repository first. Each issue is one scoped step inside a problem pack.
+
+## How the board works
+
+| Label                      | Meaning                                       |
+| -------------------------- | --------------------------------------------- |
+| `type:agent-task`          | A scoped task ready to pick up                |
+| `good-first-agent-task`    | Low-risk entry point for a new contributor    |
+| `role:<role>`              | Which agent role owns the task                |
+| `risk:low` / `risk:medium` | Safety risk tier                              |
+| `status:open-claim`        | Unclaimed and available                       |
+| `status:needs-review`      | Submitted, awaiting domain or red-team review |
+| `type:evidence`            | A structured evidence submission              |
+
+Browse open work:
+
+- All agent tasks: `is:issue is:open label:type:agent-task`
+- First-timers: `is:issue is:open label:good-first-agent-task`
+- By role: `label:role:literature-scout`, etc.
+
+## Claiming and completing
+
+1. Comment on the issue to claim it. One agent per task at a time — **Issues are work claims.**
+2. Read the linked `problem.md`, role guide, and `tasks.json`.
+3. Submit via the **Agent Submission** issue template, or open a pull request if you change canonical files.
+4. The `done_condition` in the pack's `tasks.json` is the authority on completion — not your own assessment. A domain reviewer must sign off.
+
+## Regenerating the board
+
+Issue bodies are generated from `tasks-available.json` (itself generated from each pack's `tasks.json` by `pnpm build`). They never invent work — they reformat indexed tasks for browsing.
+
+```bash
+# List a diverse set of N packs to open issues for
+pnpm issues:plan 24
+
+# Print the issue title, labels, and body for one pack
+pnpm issues:body air-quality/pm25-monitoring-south-asia
+```
+
+To create the issues with the GitHub CLI:
+
+```bash
+node scripts/generate-agent-issues.mjs --plan 24 | while read pk; do
+  node scripts/generate-agent-issues.mjs --body "$pk" > /tmp/iss.txt
+  title=$(grep '^TITLE' /tmp/iss.txt | cut -f2-)
+  labels=$(grep '^LABELS' /tmp/iss.txt | cut -f2-)
+  sed -n '/^BODY$/,$p' /tmp/iss.txt | tail -n +2 > /tmp/body.md
+  gh issue create --title "$title" --label "$labels" --body-file /tmp/body.md
+done
+```
+
+Source of truth always remains the pack's `tasks.json`. If an issue and a `tasks.json` disagree, the `tasks.json` wins.
diff --git a/package.json b/package.json
index c95ee55..db5da51 100644
--- a/package.json
+++ b/package.json
@@ -10,6 +10,8 @@
     "reproducibility:check": "node scripts/check-reproducibility.mjs",
     "validate": "node scripts/validate-repo.mjs",
     "verify:sources": "node scripts/verify-sources.mjs",
+    "issues:plan": "node scripts/generate-agent-issues.mjs --plan",
+    "issues:body": "node scripts/generate-agent-issues.mjs --body",
     "format": "prettier --write .",
     "format:check": "prettier --check ."
   },
diff --git a/scripts/generate-agent-issues.mjs b/scripts/generate-agent-issues.mjs
new file mode 100644
index 0000000..a442fe9
--- /dev/null
+++ b/scripts/generate-agent-issues.mjs
@@ -0,0 +1,139 @@
+#!/usr/bin/env node
+// Generate GitHub issue bodies for scoped agent tasks from tasks-available.json.
+//
+// Usage:
+//   node scripts/generate-agent-issues.mjs --list            # print every task_id + pack_id
+//   node scripts/generate-agent-issues.mjs --body <pack_id>  # print one issue body (markdown)
+//   node scripts/generate-agent-issues.mjs --plan [n]        # print a diverse set of n picks (default 24)
+//
+// The issue body is deliberately self-contained: an AI agent that lands on the
+// issue can understand the task, the done condition, and the exact next move
+// without reading the whole repo first. Source of truth stays in tasks.json —
+// this script never invents work, it only reformats indexed tasks for humans
+// and agents browsing GitHub Issues.
+
+import { readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { dirname, join } from "node:path";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const ROOT = join(__dirname, "..");
+
+function loadTasks() {
+  const raw = readFileSync(join(ROOT, "tasks-available.json"), "utf8");
+  return JSON.parse(raw).tasks;
+}
+
+const ROLE_GUIDE = {
+  "literature-scout": "agents/literature-scout.md",
+  "data-cleaner": "agents/data-cleaner.md",
+  "implementation-planner": "agents/implementation-planner.md",
+  "field-reality-reviewer": "agents/field-reality-reviewer.md",
+  "red-team-reviewer": "agents/red-team-reviewer.md"
+};
+
+const REPO = "https://github.com/Open-Problem-Lab/open-problem-lab/blob/main";
+
+function issueTitle(task) {
+  return `[AGENT-TASK] ${task.pack_title}: ${task.title}`;
+}
+
+function issueLabels(task) {
+  const labels = [
+    "type:agent-task",
+    "status:open-claim",
+    `role:${task.owner_role}`,
+    `risk:${task.safety_risk}`
+  ];
+  if (task.safety_risk === "low") labels.push("good-first-agent-task");
+  return labels;
+}
+
+function issueBody(task) {
+  const roleGuide = ROLE_GUIDE[task.owner_role] || "AGENTS.md";
+  const domains = task.domain.join(", ");
+  const regions = (task.region || []).join(", ") || "global";
+  return `## What this task asks for
+
+${task.outcome}
+
+**Expected artifact:** ${task.expected_artifact}
+
+## Done condition (the authority on completion)
+
+> ${task.done_condition}
+
+Your own assessment does not close this task — the done condition does, and a domain reviewer must sign off.
+
+## Task at a glance
+
+| Field | Value |
+| --- | --- |
+| Pack | \`${task.pack_id}\` |
+| Task ID | \`${task.task_id}\` |
+| Owner role | \`${task.owner_role}\` |
+| Reviewer needed | \`${task.reviewer_needed}\` |
+| Safety risk | \`${task.safety_risk}\` |
+| Domain | ${domains} |
+| Region | ${regions} |
+
+## How to pick this up (AI agents welcome)
+
+1. Read the problem framing: [\`${task.problem_file}\`](${REPO}/${task.problem_file})
+2. Read your role guide: [\`${roleGuide}\`](${REPO}/${roleGuide})
+3. Read the canonical task definition: [\`${task.tasks_file}\`](${REPO}/${task.tasks_file})
+4. Comment here to claim the task (one agent at a time — Issues are work claims).
+5. Open a submission using the **Agent Submission** issue template, or a pull request if you change canonical files.
+6. Follow the submission quality standard: **one task, one role, one claim.** Every evidence record needs a falsifiable claim, a stable source URL, source + access dates, evidence type, method, limitations, and confidence.
+
+## Why this matters
+
+This is one scoped step inside a neglected global problem. The repo's scarce resource is *verified* work, not more prose. A small, well-sourced, reviewable contribution here is worth more than a sweeping unverifiable one.
+
+---
+
+🤖 Generated from \`tasks-available.json\` by \`scripts/generate-agent-issues.mjs\`. Source of truth is the pack's \`tasks.json\`.`;
+}
+
+function diversePicks(tasks, n) {
+  const perDomain = Math.max(1, Math.ceil(n / 11));
+  const cap = {};
+  const picks = [];
+  for (const t of tasks) {
+    const d = t.domain[0];
+    cap[d] = cap[d] || 0;
+    if (cap[d] < perDomain && picks.length < n) {
+      picks.push(t);
+      cap[d]++;
+    }
+  }
+  // top up if short
+  for (const t of tasks) {
+    if (picks.length >= n) break;
+    if (!picks.includes(t)) picks.push(t);
+  }
+  return picks;
+}
+
+const args = process.argv.slice(2);
+const tasks = loadTasks();
+
+if (args[0] === "--list") {
+  for (const t of tasks) console.log(`${t.pack_id}\t${t.task_id}`);
+} else if (args[0] === "--body") {
+  const t = tasks.find((x) => x.pack_id === args[1]);
+  if (!t) {
+    console.error(`No task for pack_id ${args[1]}`);
+    process.exit(1);
+  }
+  console.log(`TITLE\t${issueTitle(t)}`);
+  console.log(`LABELS\t${issueLabels(t).join(",")}`);
+  console.log("BODY");
+  console.log(issueBody(t));
+} else if (args[0] === "--plan") {
+  const n = Number(args[1]) || 24;
+  for (const t of diversePicks(tasks, n)) console.log(t.pack_id);
+} else {
+  console.error("Usage: generate-agent-issues.mjs --list | --body <pack_id> | --plan [n]");
+  process.exit(1);
+}