Product2 publishers3 min readPublished
Autoheal's agent-fixing agents still need an engineer to approve every change
Autoheal raised a $7.9 million seed round for a platform that scores a company's AI agents and repairs the weak ones. Its customer evidence so far covers incident investigation, so teams buying it for vulnerability fixes are buying on the pitch alone.
The Product Desk · Product desk
What happened
- Innovation Endeavors led the round and Harpinder Singh joins Autoheal's board, with Emergent Ventures, U&I Ventures, Darkmode Ventures, Batch Ventures and Param Hansa Values also investing.
- An Evaluator agent scores each worker agent's run using review comments, CI failures and the incidents that agent caused.
- Every change the Healer agent proposes to an agent's prompts, tools, skills or model is version-controlled in git and needs an engineer's approval.
- Choudhury said the next goal is reinforcement learning on customers' private engineering data to build company-specific small language models.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost The hours spent reviewing each proposed agent change come out of the same engineering capacity Autoheal says repetitive work already consumes, and the buyer's platform team pays them.
- capability With agent behavior stored in git, a platform team can diff, review and revert a prompt or model change using the tooling it already uses for code.
- exposure One platform connected to repos, CI/CD and observability concentrates access a team would otherwise grant tool by tool in a single vendor's software, even when it runs in the customer's cloud.
- precedent Grading agents on CI failures and incidents they caused gives buyers an outcome-based standard they can ask any agent vendor to report against.
At Nomura, the work Autoheal took over starts in an alert queue. "Our production operations teams spend valuable time triaging alerts and managing incidents, while also pulling engineers away from their software development activities," said Sameer Jain, the bank's CIO for wholesale [5]. He gave speed and location as the reasons it fit: "Autoheal gives us a platform that takes investigation timelines down from hours to minutes. The fact that it runs entirely within our own cloud, in compliance with our controls, made it a natural fit for how we operate," he said [6].
The pitch around that quote is wider. Autoheal says off-the-shelf point agents failed to deliver at Nomura and AvidXchange [3]. It also says repetitive work such as incident response and vulnerability remediation takes more than a third of an engineering team's capacity [4]. Both are the company's statements. Jain's words are about alert triage and where the software runs [6]. According to SiliconANGLE, customers including AvidXchange and Empiric Earth say they cut incident resolution times and saved thousands of hours of engineering work [13].
In a "self-improving software factory" [1], a platform engineer opens a pull request the Healer wrote against an agent's configuration, already checked against historical benchmarks for regressions [8], and decides whether to merge it. Autoheal says engineers widen agent autonomy as agents prove reliable [17]. The Evaluator's evidence for that reliability arrives downstream of the work. A caused incident, the most serious of its signals, can only be counted once code is in production [7].
Sid Choudhury, Autoheal's co-founder and CEO, frames the problem as upkeep. "Our experience taught us that while building the first version of an AI agent is easy, scaling it consistently across the enterprise SDLC is the real challenge," he said [10]. He said the product lets platform engineers "immediately step into the role of AI engineers and accelerate ROI, without spending a year building the underlying infrastructure" [11]. The idea came at Harness, where he and his co-founders saw the need to "manage agents as code, overseen by continuously learning meta-agents," he said [15]. Autoheal says its goal is lower cost per successful task, along with higher accuracy and faster execution [18]. The announcement does not include a price, a deployment time, or any customer's cost per task.
I'd sort a team on two axes: how many agents already work its delivery pipeline, and whether production access has to stay inside its own cloud. One agent, data free to leave: a point tool is simpler to evaluate, and a shared context graph has little to share [12]. One agent, data kept in-house: self-hosting is the reason Jain gave [6], and the Healer's queue stays short. Several agents, data free to leave: the scaling problem Choudhury describes applies [10], and the open question is who owns the reviews. Several agents, data kept in-house: Autoheal's private-cloud deployment was built for this buyer [12].
In every quadrant, I'd score a pilot on cost per successful task with approval hours counted in. Every behavior change the Healer proposes waits for an engineer [9].
What to watch
- A published customer result for vulnerability remediation or coding-agent quality, beyond faster incident investigation.
- Autoheal's pricing, and whether it is set per agent, per task or per successful task.
- How often engineers at Nomura or AvidXchange accept the Healer's pull requests, and whether any agent is allowed to change without approval.