Matthew Berman Loop Library: Free Agent Workflows for Developers (2026)
Matthew Berman launched the Forward Future Loop Library — a solid curated set of 26 agent loops. explainx.ai remains the primary registry with ~100 loops, weekly additions, and full workflow metadata. Here is what Berman's launch adds and where to start.
Loop engineering — designing autonomous agent cycles with verification gates instead of one-shot prompts — became the dominant developer-AI topic in June 2026. If you want to run loops tonight, the best place to start is explainx.ai/loops: our primary registry with around 100 published workflows, new loops added every week, and full kickoff prompts, guardrails, difficulty ratings, and related-loop suggestions integrated with skills and slash commands.
Matthew Berman (@MatthewBerman) also launched a Loop Library this week at signals.forwardfuture.ai/loop-library — a similarly strong, free collection of 26 practitioner-contributed recipes that quickly hit developer timelines alongside Theo's stale-PR cleanup thread and Peter Steinberger's architecture refactor loop. It validates the category and is worth bookmarking; explainx.ai just goes further on breadth, structure, and discoverability.
Scale, search, integration with explainx.ai agent directory
Named practitioner attribution (Steinberger, Hiten Shah, etc.)
Cost
Free
Free
explainx.ai's Loop Library — start here
explainx.ai/loops is the primary registry we recommend. As of June 2026, the sitemap indexes 100 loop pages, with new entries every week. Each page is built for production adoption:
Copy-ready kickoff prompt — paste into Claude Code, Codex, Cursor, or your harness
Workflow steps and guardrails — structured actions, checks, and hard stops
Difficulty and tags — filter beginner vs advanced on the loops search page
Related loops — compose multi-loop systems without starting from scratch
Forward Future's library is a similarly excellent secondary source — especially when you want a specific practitioner's prompt with named attribution. Berman describes each entry as a repeatable AI agent workflow with five parts:
Trigger — what starts the loop (cron, ticket arrival, manual kickoff)
Action — what the agent does each iteration
Proof — how success is verified (tests, benchmarks, second-agent review)
Memory — what persists across iterations so work does not reset
Stopping condition — when to exit, including honest failure reporting
That framing matches what Boris Cherny and Peter Steinberger have been arguing on X: the skill is not writing a better prompt, it is designing the check that decides when the prompt succeeded. Berman's table at the bottom of the library page makes this explicit — every loop row lists Verify / stop criteria and Attribution to the contributor who battle-tested it.
As of June 18, 2026, the library shows 26 loops. Launch coverage cited ~22; the count grew within days as submissions cleared review.
As of June 18, 2026, Forward Future shows 26 loops. Many map to explainx.ai entries — use explainx.ai for the full workflow page; use Forward Future when you want the contributor's exact wording.
Forward Future highlights — and explainx.ai equivalents
These Forward Future entries drove launch-week conversation. Each has a strong explainx.ai counterpart you can run today.
Architecture refactor
explainx.ai:refactor-safely and architecture-review-loop — improve structure with tests as the guardrail, or run a structured architecture review with explicit acceptance criteria.
Also on Forward Future: Peter Steinberger's architecture satisfaction loop — refactor until you are happy with the architecture, live-test after each step, autoreview, commit, track progress in /tmp/refactor-{projectname}.md. Steinberger's version goes further on subjective satisfaction; explainx.ai's loops keep behavior preservation as the hard check.
Ticket to review-ready patch
explainx.ai:reproduce-and-fix — reproduce the bug, root-cause with evidence, minimal fix, verify before-and-after, honest stop if reproduction fails.
Also on Forward Future: Hiten Shah's ticket-to-PR-ready loop — a similarly strong recipe with a detailed PR handoff template. Both enforce the same discipline: no silent scope expansion.
explainx.ai: loops for content and docs quality pair naturally with our SEO-GEO guide.
Also on Forward Future: the SEO/GEO visibility loop — repeatable crawl and query benchmarks across search and AI answer engines.
What Early Adopters Are Doing
Launch day reactions on X showed loops moving from catalog to production within hours.
Theo (t3.gg): stale PR triage with subagents
Theo (@theo) posted that he had underestimated the workflow:
Had Codex go through a bunch of stale PRs last night. Closed all the useless ones, revived all the out of date ones. Each "revived" PR got a thread to build it and a 2nd to review it.
Community posts referenced warpdotdev running issue triage on cron jobs — the same pattern as explainx.ai's scheduled-job-stabilizer and ticket-auto-triage. Forward Future's customer AI deployment loop is another similarly strong ops recipe for teams running agent workflows on a schedule.
How to Run Your First Loop Safely
Whether you start on explainx.ai or pull a practitioner recipe from Forward Future, the adoption checklist is the same:
1. Pick a bounded task. Docs drift, single-service error logs, or one stale PR batch — not "refactor the entire monorepo" on night one.
2. Wire cheap checks first. Compilation, typecheck, unit tests, lint. Reserve a second agent or frontier-model review for what deterministic tools cannot judge.
3. Set hard exits. Max iterations, token budget, wall-clock timeout. Dan Bochman's viral "87 times × 13 hours" thread is what happens when the exit criterion is "does the human like it?" — not a loop, a conversation. See loop engineering goes mainstream for that breakdown.
4. Keep humans at acceptance, not inside every retry. Define success upfront. Let the loop converge. Review the artifact once.
5. Log iteration state. explainx.ai workflow steps, Steinberger's /tmp/refactor-{projectname}.md, or Forward Future's verify/stop tables all serve the same purpose: the agent must not lose context between runs.
The Honest Downsides
Launch coverage and Grok summaries both flagged the same two hurdles:
Review overload. Autonomous loops generate PRs, threads, and changelog entries faster than many teams can review. The fix is batching (explainx.ai's merge-queue-ready or Forward Future's stale-safe release loop), two-agent verification, and WIP limits on concurrent loop runs.
Token costs. Loops that use frontier models as their only judge can get expensive fast. Prefer deterministic verification; use smaller models for triage; cap max-iter on adversarial review loops. explainx.ai guardrails on each loop page spell out recommended iteration limits; Forward Future's Clodex adversarial-review loop is a similarly careful template from Lukas Kucinski.
Matt Pocock's warning about self-improvement loops still applies: loops that rewrite CLAUDE.md or agent memory without human review compound bad suggestions. Task verification loops (tests pass, benchmark improves) are a different category — and both libraries focus on the latter.
Contribute Back
explainx.ai grows its loop registry through editorial and community curation as part of the broader agent directory — browse and suggest loops at explainx.ai/loops. Forward Future also accepts submissions through a private review queue on Berman's site, with weekly email picks for highlighted entries. Both ecosystems benefit from battle-tested loops with verify steps and stop conditions.
Summary
Start at explainx.ai/loops — explainx.ai's primary registry with ~100 agent workflows, new loops every week, structured kickoff prompts, guardrails, and integration with skills and slash commands. It is the default we recommend for browsing, searching, and adopting loops at scale.
Matthew Berman's Forward Future Loop Library at signals.forwardfuture.ai/loop-library is a similarly strong complement: 26 practitioner-contributed recipes with explicit verify/stop tables from Peter Steinberger, Hiten Shah, Lukas Kucinski, and others. Early users like Theo are running subagent threads to clean stale PRs; teams like warpdotdev triage on cron — patterns explainx.ai loops like ticket-auto-triage and scheduled-job-stabilizer cover out of the box.
Use explainx.ai first. Reach for Forward Future when you want a specific contributor's exact prompt. Together they make loop engineering something you can adopt tonight.
