When a production system at a large bank breaks in the middle of the night, the engineer who gets paged usually spends the first hour doing detective work across dashboards, logs, Slack threads and recent code changes. At a median cost of $2 million an hour for a high-impact outage, that hour of searching is where most of the money disappears, because nothing can be fixed until someone knows what actually broke. AI coding tools have made the hour harder to protect, because enterprises spent $4 billion on them in 2025, about seven times what they spent the year before, while developers pushed nearly one billion commits to GitHub over the same period, which means more changes reaching production every day and more places for something to go wrong.Autoheal, a San Francisco company whose AI agents already handle that detective work inside Nomura, has raised $7.9 million to build what it calls a self-improving software factory, a platform where agents take on the repetitive jobs that follow the code, from incident response and security patching to keeping AI coding bills under control. The healing in Autoheal runs in two directions, since its worker agents fix problems in a company's production systems while a second layer of agents watches the first and repairs them whenever their performance slips, which makes it one of the most complete attempts so far at enterprise software that keeps itself healthy.https://www.youtube.com/watch?v=f3nj6cYKuBA&embedable=trueThe RoundThe round was led by Innovation Endeavors, the Palo Alto early-stage firm that backs founders in intelligent software and computing infrastructure, with its partner Harpinder Singh taking a seat on Autoheal's board. Singh came to venture capital as an operator, having co-founded and run the ecommerce data company Slice Technologies until Rakuten acquired it, before sitting on the board of the identity-security company Authomize, which is now part of Delinea, so he arrives with experience of both enterprise data and security buyers. Emergent Ventures, U&I Ventures, Darkmode Ventures, Batch Ventures and Param Hansa Values also joined the round.The customer list is what sets this seed apart, because Autoheal only emerged from stealth this year and is already running in production at Nomura, the payments company AvidXchange and Empiric Earth, three environments where security reviews routinely stop young AI vendors at the door. Autoheal gets through those reviews by running a three-week proof of value in which one of its engineers sits with the customer's team and measures success on the customer's own numbers for incidents, time to resolve, change lead time and cloud and AI spend. The new capital will go into repeating that motion across more enterprise engineering teams while funding the company's work on models trained on each customer's private data.AI Coding and the Post-Coding WorkloadWriting code has become the first job that enterprises pay generative AI to do at real scale. Menlo Ventures, which surveyed 495 enterprise AI buyers for its 2025 State of Generative AI in the Enterprise report, found that total corporate spending on generative AI jumped from $1.7 billion in 2023 to $37 billion in 2025, with coding now the largest category of spending by any single department, which works out to roughly 11 cents of every enterprise AI dollar going into tools that write software. Gartner expects 90% of enterprise software engineers to use AI code assistants by 2028, up from under 14% in early 2024, so the adoption curve in the chart below still has most of its climb ahead of it.A useful way to picture what that spending does is a factory that doubles the speed of its assembly line without hiring more inspectors or repair crews, because it ships more finished goods and also more faulty ones, which is close to what Google's 2025 DORA report measured across nearly 5,000 technology professionals. DORA found that 90% of them now use AI at work, up from roughly three in four a year earlier, with heavier AI use linked to teams shipping more software but also to more failed changes, more rework and longer recovery times, a pattern the researchers put down to testing, review and operations not keeping pace with the code arriving from upstream.Autoheal estimates that repetitive work such as incident response and vulnerability fixing already takes more than a third of an engineering team's capacity, which for a 900-person engineering organisation would mean the equivalent of about 300 engineers spending their days on triage and firefighting rather than on the product.Security and Reliability PressureSecurity teams feel the extra code first, because every new line is another place for a flaw to hide. A record 48,185 CVEs were published in 2025, up 20.6% on 2024 and the highest annual total since the CVE programme began in 1999, while the first half of 2026 added 35,364 more, which is 49.5% above the same stretch of 2025 and puts this year on course for more than 71,000, or roughly 195 newly catalogued flaws every single day. NIST said in April that CVE submissions had risen 263% between 2020 and 2025 and that its analysts would enrich only the highest-priority entries from now on because they could not keep up, while Veracode's tests of more than 100 large language models found that 45% of AI-generated code samples introduced a known security flaw, so some of the new volume is being written by the same tools that are speeding everything else up.Speed matters even more than volume here, because a flaw only costs money once someone uses it. Mandiant's research on time-to-exploit shows the average gap between a vulnerability going public and the first attack using it shrinking from 63 days in 2018 and 2019 to five days in 2023, while later analysis of the same series puts it below zero for 2024 and 2025, meaning the typical exploited flaw was already under attack before the world knew it existed. Most organisations still need somewhere between 55 and 61 days to fix a critical flaw, according to Edgescan data compiled by ProjectDiscovery. Because the two clocks were roughly level in 2018, the whole of that two-month repair window has now become time in which attackers hold the advantage, which is why the triage, patching and testing work that sits in human queues is moving to software that runs around the clock.Outages turn the same pressure straight into lost revenue, which is where the case for faster recovery becomes easiest to price. New Relic's 2025 Observability Forecast, a survey of more than 1,700 IT and engineering leaders in 23 countries, puts the median cost of a high-impact outage at $2 million an hour and the median annual bill for such outages at $76 million per company, which implies the typical large business loses about 38 hours a year to serious downtime. If every one of those hours shrank by the 87.5% that Nomura recorded when its resolution time fell from two hours to 15 minutes, the modelled annual bill would drop from $76 million to about $9.5 million. Real outages rarely compress that neatly, yet the arithmetic shows why engineering leaders now treat time to resolve as a financial number rather than an operational one.How the Self-Healing Loop WorksAutoheal begins by plugging into the tools an engineering organisation already runs, including its coding agents, code repositories, build and deployment pipelines, monitoring systems, cloud accounts and ticket trackers, then turns everything it learns from them into a shared map of how the company's systems fit together. Worker agents draw on that map to do specific jobs such as investigating an alert, patching a vulnerability or checking whether a release is safe to ship, while any team can create a new agent by describing the job in plain language, after which Autoheal turns the description into a workflow, tests it against the customer's own evaluations and rolls it out. Because every agent shares the same memory, a fix that one engineer teaches an agent on a Monday becomes something every other team in the company can use by Tuesday.The harder problem that Autoheal was designed around is that AI agents quietly get worse over time as the systems around them keep changing. An agent that gave accurate answers in March can be wrong by June because the codebase it reads has changed, a service it depends on has been renamed or the model underneath it has been swapped for a newer one, yet most companies only find out when an engineer notices the answers slipping. Autoheal handles this with two background agents, an Evaluator that scores every worker run against what actually happened afterwards, so a coding agent is judged on the review comments, failed builds and production incidents its changes go on to cause, plus a Healer that takes the lowest-scoring agents and opens pull requests to improve their instructions, tools or choice of model. The Healer replays each proposed fix against past runs to make sure nothing else breaks before an engineer sees it. Every change is stored in git and needs a human sign-off, so the system improves itself while engineers keep the final word and hand over more autonomy only as each agent proves reliable.Cost gets the same treatment, since AI coding tools bill by usage and a busy engineering organisation can run up large token bills without noticing. Autoheal routes each task to the cheapest model that can do it well, batches calls and tunes caching automatically, while dashboards show cost, speed and accuracy for every team and every agent, which lets a platform team manage toward cost per successful task, a far more useful number than raw spend. On security, the platform can run as a cloud service, as a hybrid or fully air-gapped inside the customer's own cloud with pre-approved models, with each agent working in an isolated environment on short-lived, narrowly scoped credentials, while the company lists SOC 2 Type II and ISO 27001 certification with zero data retention.Customer DeploymentsNomura shows what the product can do under the tightest conditions, since a global investment bank running regulated production systems cut its mean time to resolve incidents from two hours to 15 minutes after deploying Autoheal. The bank's wholesale technology leadership has singled out the fact that the platform runs entirely inside Nomura's own cloud and within its existing controls as the reason it fit the way the bank works, which speaks directly to the first question any financial institution asks before giving an AI agent access to production.AvidXchange, which automates accounts payable and payments for mid-sized businesses, uses Autoheal for production incident response and saves thousands of engineering hours every month that now go back into building features. Its engineering leadership plans to move the platform earlier in the development process next, which is the natural expansion path for a product that starts in operations. Empiric Earth uses Autoheal to speed up troubleshooting across a complicated environment while cutting what it spends on monitoring software, while the risk-decisioning company Oscilar uses it to triage customer support tickets by pulling together data from Grafana, Slack, ClickHouse, its product documentation and Pylon so that engineers start each case with a likely root cause already in hand.Founder ThesisChief executive Sid Choudhury has spent his career on the tools engineers use to keep software running, which gives Autoheal an unusually direct line into the problem it is solving. He was the first product manager at AppDynamics, the application-monitoring company that Cisco later acquired, then led product at the database company Yugabyte before becoming senior vice president and general manager at Harness, the software delivery company that is on track to pass $250 million in annual recurring revenue at a $5.5 billion valuation. Chief technology officer Utkarsh Ohm led AI and machine learning engineering at ThoughtSpot after joining HyperTrack as a founding engineer, while chief development officer Puneet Saraswat was vice president of engineering at Harness after managing engineering teams at Microsoft Azure.Having helped scale Harness past $200 million in annual recurring revenue, the three founders saw a pattern forming across enterprise engineering teams, where building a first AI agent had become easy while running dozens of them safely across many teams was eating months of work and large token budgets. Their conclusion was that agents should be managed the way good platform teams already manage infrastructure, as code in version control that is supervised by meta-agents which keep learning, so that a company ends up with one governed fleet rather than a sprawl of disconnected point tools. The pitch to platform engineers is that they can step into the role of AI engineers immediately and show a return within weeks, without first spending a year building the evaluation, governance and context layers that every serious agent deployment eventually needs.Investor ThesisInnovation Endeavors is backing a shift in the question enterprises ask about AI agents, which has moved from whether a single agent can do a task to how hundreds of them can run safely and cheaply across a whole engineering organisation. Singh describes Autoheal as the infrastructure layer that makes that second question answerable, seeing an opportunity far larger than any one agent or workflow because the platform gives engineering teams a repeatable way to put specialised AI to work at every stage of building and running software.The buyers for that layer already exist in large numbers, since DORA found that 90% of organisations now run at least one internal developer platform, usually looked after by a dedicated platform team that decides which tools every other engineer uses. Harness has shown how much value sits in this part of the market by building a business now valued at $5.5 billion around what it calls "AI for everything after code." Autoheal arrives in that market with founders who helped build that company, a design that fits regulated industries from the first day and paying customers in banking and payments, which are the sectors where trust in AI takes longest to earn and is worth the most once it is earned.The Future is Super Intelligence: The Sovereign Intelligence RoadmapAutoheal's next step is what it calls sovereign intelligence, which means using reinforcement learning on each customer's private engineering data to train small language models that belong to that company and never leave its security boundary. Frontier models know an enormous amount about the outside world yet very little about how one bank's payment services depend on each other, so a smaller model trained on a company's own incidents, code and runbooks can power every agent in its factory at lower cost while understanding the business far better than a general model could. The company intends to carry the same design beyond software into data engineering and security engineering over time, since both are full of the repetitive, high-stakes operational work that its agents already handle.AI will keep writing a larger share of the world's software every quarter, with each of those changes adding to the incidents, patches and token bills that someone has to manage once the code is live, so the companies that build the tools for running software after it ships are sitting in one of the fastest-growing parts of enterprise technology. Autoheal expects every large enterprise to operate its own software factory staffed by specialised agents, with Autoheal as the platform those companies use to build, govern and keep improving them. With $7.9 million in fresh capital, an experienced operator from Innovation Endeavors on its board and results at Nomura and AvidXchange already on the record, it goes after that market having shown that software which heals itself can pass the test set by some of the most demanding engineering organisations in finance.Don’t forget to like and share the story!Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR.