The moment arrived quietly for most engineering teams: AI agents graduated from demos to deployment, earning write access to repositories, infrastructure dashboards, customer support queues. They merge pull requests now. File tickets. Drop messages into Slack channels at 3 a.m.
And when they fail—and they do fail—the discovery process is often, well, manual.
Enter Archal, a two-person operation that emerged from Y Combinator's Summer 2026 cohort with a specific thesis: production agents need their own reliability loop, one that catches failures, reconstructs them in isolation, patches the code, and proves the fix works before any human reviews a line. The founders, Noah Song and Aidan Tiruvan, are betting that teams running stateful agents—those interacting with live systems like Stripe, GitHub, or Linear—can't rely on post-mortem debugging much longer.
Their platform, now in Early Access, doesn't just log errors. It replays them against what the company calls "clones": lightweight, stateful copies of real services that hold state, enforce referential integrity, and even mimic the occasional cryptic error message you'd see from a real API. When a failure surfaces, Archal spins up another coding agent to write a fix, opens a pull request in the team's repo, and runs the failed scenario again. Only if it passes does the PR land in a reviewer's queue.
Nothing merges without proof.
The Loop: Catch, Recreate, Fix, Prove, Remember
It's a five-stage cycle, and the sequencing matters. Archal pulls production traces from an agent's existing observability tooling—or its own SDK if tracing infrastructure doesn't exist yet. Failures get flagged, graded, then recreated inside isolated containers running those service clones. A coding agent steps in to patch the agent harness: the prompts, tool wiring, glue logic that connects an LLM to the outside world. The proposed fix runs against the same failed scenario. If satisfaction scores—a 0-to-100 metric the platform calculates—clear the team's threshold, the fix moves forward. If not, the loop retries.
Fixed failures become stored evaluations, a library of "this used to break, it shouldn't again" test cases. Whether that actually prevents regressions at scale remains to be seen in practice—the company hasn't shared public case studies or customer data yet—but the intent is clear.
The clone inventory has grown to 22 services. Four are public: GitHub, Linear, Slack, Supabase. Eighteen more sit in preview, including Stripe, Ramp, Jira, Discord, Google Workspace, Datadog, HubSpot, Sentry, Webflow. Some, like Ramp, are Model Context Protocol-only for now, exposing tools for card management, transaction lookups, reimbursement flows. Others, like Slack, implement the full MCP spec with message posting, threading, reactions, even pre-seeded workspaces labeled "engineering-team" or "incident-active."
The clones run in isolated containers. They never phone home to real services. Agent API keys stay local, and according to Archal's security documentation, the company's servers never touch credentials for OpenAI, Anthropic, or other LLM providers. It's a deliberate architecture—perhaps a necessity, given how twitchy enterprises get about letting third parties near production tokens.
CI Pipelines and the Gating Question
Teams can wire Archal into GitHub Actions or GitLab CI. Set a workspace API key, define a satisfaction threshold, and the build fails if agent performance dips. The CLI spits out that 0-to-100 score per run. Commands like archal inspect and archal export let engineers bundle tool calls and state snapshots, with optional anonymization for sensitive data.
The quickstart is lean: Node.js 20 or higher, three commands (npx archal init, archal login, connect the harness), and scenarios written in markdown with setup instructions, a prompt, success criteria. The platform supports inline tasks and repeated runs to check consistency, though how teams define "consistency" for a stochastic agent is its own rabbit hole.
Archal integrates with 14 agent frameworks and SDKs—OpenClaw, LangGraph, LlamaIndex, AutoGen, Anthropic SDK, CrewAI, Mastra, OpenAI Agents SDK among them. A feature the docs call "route-mode" can redirect SDK traffic to clones without code changes, though that capability isn't live for all services yet.
Why Now? The Timing Question
GitHub's July 10 announcement of agentic autofix for code scanning alerts—a public preview that lets organizations with Code Security and Copilot Cloud Agent automatically remediate static analysis findings and open PRs—was a signal flare. Major platforms are embracing agent-authored code changes at scale, AI Credits be damned. The feature targets SAST alerts specifically, but the precedent is what matters: agents fixing code, unsupervised, in production repos.
Elsewhere, the ecosystem is thickening. Arga Labs launched real-world sandboxes with API twins for Stripe, Slack, Google Drive. Raucle built verifiable audit records and integrated with Microsoft's Agent Governance Toolkit in May. Archron launched a control tower that pre-authorizes agent actions before they execute. Concordium's Agent Registry claimed 1,131 verified AI agents by mid-July, seven weeks post-launch. Research papers started appearing in June and July analyzing agent-authored PR frequency, conflict rates, the shortcomings of naive evaluation metrics.
In other words, teams are handing agents more permissions, and the guardrails—or at least the monitoring layer—are scrambling to catch up.
What's Still Under Wraps
Archal hasn't disclosed pricing, customer logos, or funding details beyond standard Y Combinator participation. The YC directory lists the team size as two, though that figure could lag real-time headcount. Early Access users can book a call through the site. A Reddit thread titled "Today we're excited to introduce Archal (YC S26)" appeared on May 27, but the post content has since been removed—make of that what you will.
The platform's documentation was updated within the last couple of months. Telemetry is off by default, with optional PostHog analytics that teams can toggle via config or environment variables. The company describes itself as "built for stateful agents," a framing that distinguishes it from purely code-focused autofix tools or generic sandboxes. It's a narrow positioning, but perhaps deliberately so.
What This Means for Teams Deploying Agents
Archal's model is opinionated. It assumes teams are past prototyping, running agents that touch multiple services and need proof of correctness before merging changes. The clone-based replay and automated PR workflow treats the agent harness—not the LLM provider—as the artifact to patch. It requires teams to connect both their repository and their trace store upfront, which isn't trivial plumbing.
Whether that fits a given team depends on how they've structured their agent stack. Some will prefer human-in-the-loop verification at every step. Others may want guardrails that block actions before they execute, not after they fail. Archal occupies a specific point in that spectrum: post-failure, pre-merge, with a coding agent doing the remediation work.
For teams already debugging agent failures manually—or worse, retrofitting evaluations after production incidents—the loop offers a faster iteration path. The real test is whether the automated fixes hold, and whether those stored evaluations actually prevent regressions when the codebase evolves. That evidence isn't public yet.
But the product is live. Teams can try it. And the underlying question it poses is worth sitting with: if we're giving agents production credentials, who's responsible when they break? Another agent, apparently. At least in Archal's vision of how this scales.
