A solo founder shows what running a software company with AI agents really involves

One founder, dozens of agent sessions

As docspring.com reported in the founder's own account, DocSpring, which fills and generates PDFs and Word documents and collects e-signatures, now does much of its development and operations with AI agents. The founder, who started the company as FormAPI and runs it solo, says he writes far less code himself yet works harder than ever.

The payoff is a backlog that finally moves. He lists Word templates, a browser-based Page Designer, API v2, a redesign of the website and app, and a lot of e-signing work as things that shipped in recent months. Side-by-side previews, requested by a customer years ago, also made it out of the queue.

How the daily work is organized

He uses a tool called Herdr to manage 49 Claude Code sessions, each focused on one task or product area. Of those, 26 are currently active and only about a dozen work in parallel. Much of his day is reading updates and answering questions that the sessions raise. He says the constant switching left him exhausted at first, but he has adapted.

Some sessions have standing jobs:

• One works through batches of bug fixes, from investigation to review.
• One handles Sentry, the service that reports application errors, and sorts real problems from noise.
• One investigates slow pages, tracing database queries and measuring results after deployment.
• One maintains the CI pipeline, the automated checks that run on new code, including moving many jobs to his home servers.

The result, he says, is that reliability and performance work no longer has to wait while he builds something bigger.

Research that used to be out of reach

The clearest example is the effort to seal PDFs with a certificate trusted through the Adobe Approved Trust List. When he looked into it a few years ago, contracts started at $50k per year, which was too much for a small startup. Agents helped him research and contact every vendor they could find, and he now has one that charges only a few thousand dollars per year and lets him bring his own cloud HSM (a hardware-backed key store in the cloud). The certificate has not yet been issued.

He also describes Sorbet, a type checker for Ruby, which agents helped apply to the whole app. He says maintaining custom plugins for it was a project he once considered unrealistic.

Guardrails and the inbox

Changes still pass automated checks, code review and a staging environment before reaching production. He built a tool called Rack Gateway so agents can prepare server commands and deployments, then hand him the actual code or commit to approve. Sometimes five agents wait in the deploy queue at once.

Email is handled by a saved set of instructions, called a skill, that reads threads, summarizes them and proposes actions for his approval. It can forward a message to the session already working on that topic, such as a SOC 2 compliance task. He had to set a clear boundary, triage and route only, because otherwise the skill started taking on the follow-up work itself.

For about a month, he says, his inboxes, Intercom and Sentry projects have all been empty, which he could rarely manage before.

When review goes the wrong way

The most useful part of the post is a failure. A bug in a date-difference feature came from using an average year length to compute age, which could be wrong around a birthday. The fix was to count calendar years properly.

During review, an agent noticed a nearby inconsistency: a two-digit year could be read differently in the browser and in the generated PDF. Each later review round found another mismatch, and the fixes kept widening what the parser accepted, eventually including values like 07MAY24 and 7th May 24. That contradicted an earlier decision: API v2 validates dates strictly, with only a narrow allowance for two-digit years in v1.

The agent had not checked those earlier discussions, and the founder admits he had not read every approved suggestion closely. The change was already on staging when he questioned it. The extra parsing was removed before production, and only the simple calendar fix shipped. His lesson is that repeated review can carry work further in the wrong direction when nobody holds the intended behavior in mind.

What he does not claim

The founder is cautious about the bigger picture. He says none of this proves DocSpring is safe from competition, since everyone has access to the same frontier models and coding tools, and a rival could just as easily use agents for operations. Being busy, he notes, is not evidence that customers will keep choosing his product.

It is a single company's account, and the source text available to us ends mid-thought on how long this division of labor will last.