Autoheal Raises $7.9M to Help Enterprises Evaluate and Repair AI Agents

Autoheal has secured seed funding for a system in which AI agents evaluate other agents and propose repairs—but engineers decide which changes go live. The company announced a $7.9 million round led by Innovation Endeavors on September 28, 2026, to scale its platform for enterprise engineering teams.

Innovation Endeavors’ Harpinder Singh is joining Autoheal’s board. Emergent Ventures, U&I Ventures, Darkmode Ventures, Batch Ventures and Param Hansa Values also participated. The financing backs a business focused on work around software delivery: investigating production incidents, fixing security vulnerabilities and controlling AI coding costs, rather than simply generating more code.

Individual coding help versus shared operational work

In its funding announcement, Autoheal draws a distinction between novel feature development and repetitive operations. It argues that a coding agent guided by an individual engineer can suit new feature work, but operational tasks need consistent results even when that engineer is unavailable. Its proposed alternative is a shared cloud system running background agents under common rules.

Autoheal calls that system a “software factory.” It connects repositories, automated build-and-test pipelines, monitoring tools, cloud environments and issue trackers into a context graph—a linked map of engineering information. A shared registry also stores instructions, skills and memories that agents read before acting. Autoheal says it does not replace developers’ existing coding agents; it changes the context those agents consume.

A repair proposal, not an unchecked rewrite

Two specialized agents drive the improvement loop. The Evaluator scores worker-agent runs using what happens downstream: code-review comments, failed automated checks, alerts and an incident’s actual root cause. The design uses outcomes from the company’s own software workflow to judge a run, rather than treating completion of the agent’s immediate task as sufficient.

The Healer turns weak scores into proposed changes. Those can include revised instructions, a new memory, narrower tool access or a different model. It opens a pull request—a proposed change for engineers to review—and checks the revision against historical runs for regressions. Every behavior change is tracked in git, the version-control system, and requires engineer approval.

That makes “self-improving” a bounded claim: the agents propose and test adjustments, while people retain the release decision. Autoheal also describes a separate governance layer for budgets, model selection, identities and approval policies. Agents can escalate a step to engineering or operations teams when input or permission is needed.

Controls Autoheal says enterprises can keep

• Deployment: run the platform in the customer’s own cloud or in an air-gapped environment isolated from outside networks.
• Access: start with read-only permissions, apply fine-grained policies and retain an audit trail of every tool call.
• Evaluation: score agents on the customer’s own runs inside its security boundary, without using another customer’s data to shape them.

Customer testimony now, private models later

Autoheal names Nomura Bank and AvidXchange as customers. In testimony published with the announcement, Nomura’s wholesale CIO Sameer Jain said investigations fell from hours to minutes, and emphasized deployment within the bank’s cloud and controls. AvidXchange CTO Krish Shetty said incident root-cause identification fell to minutes and described plans to extend use across more of the software lifecycle.

Those statements describe customer experience, not a guarantee for another deployment. Autoheal offers a three-week proof of value in the buyer’s environment, with success measures agreed in advance. It says it measures against the customer’s baseline, including time to an evidence-supported root cause, effort to validate a vulnerability fix and cost per successful coding task.

The longer-term plan goes beyond adjusting agent context. Autoheal wants to train enterprise-specific small language models on private engineering data within each customer’s approved security boundary. Its blog describes reinforcement learning as part of that foundation. That remains a development ambition, distinct from the evaluation-and-review loop the company describes today.