Building with AI agents while keeping human authority¶
October 5, 2026 · Kishore Babu Gutta
Historical process snapshot: This article describes my founder workflow as it stood on October 5, 2026. For the current public Runtime version and supported features, see the release overview.
As a solo founder building Engineering Runtime, I use AI for research, design, implementation and review.
The challenge is deciding which findings deserve work, keeping decisions intact across sessions, and knowing what has actually been proved.
A useful conversation can produce several ideas. A code review can uncover a genuine defect. A new provider feature can change an assumption. Without a process, each becomes another disconnected task.
I have been building a workflow that gives those inputs a place to go.
Agents can discover, propose, implement and verify. The founder retains authority over product direction, release scope and publication.
This article describes that development workflow. It is not a claim that Engineering Runtime itself orchestrates these agents or enforces all of these repository rules.
This is a historical description of my founder workflow, not a feature provided by ENGRT or evidence of a customer-validated process.
Turn findings into evidence, not automatic commitments¶
Research, code review, dogfooding and user feedback all generate useful signals. I record meaningful findings with their source, context and potential impact, then decide whether they justify a change.
AI can help compare options, identify reproducible defects and draft bounded proposals. A finding is not automatically a roadmap item, and a proposal is not automatically approved work.
Keep durable, reviewable context¶
I maintain versioned records for product decisions, potential work, implementation and release evidence. The key distinction is between what might be useful and what has been authorized.
This makes handoffs across contributors and AI sessions more reliable without treating an agent's conversational memory as a source of truth. The specific repository layout and automation arrangements are implementation choices, not requirements of ENGRT.
Scope changes deliberately¶
Before authorizing implementation, I look for duplicated work, unclear dependencies, missing exclusions and evidence that would prove the intended outcome. AI may help develop alternatives and challenge assumptions, but product scope, release readiness and publication are human decisions.
A proposed change should state the user outcome, relevant constraints and acceptance evidence. It should not expand into unrelated work simply because an agent suggests improvements.
Human authority needs explicit boundaries¶
“Human in the loop” is not specific enough to tell an agent when it should stop.
| Agent can prepare or perform within its assigned authority | Human decision retained |
|---|---|
| Research findings and proposal alternatives | Product direction and architectural changes |
| Recommended release membership and proposed splits | Acceptance of scope and allocation changes |
| Review of proposed release scope against current evidence | Which work is approved for the release |
| Implementation plan and reviewable changes | Proposal approval and release activation |
| Test results and qualification report | Scope exceptions and publication authorization |
| Reviewing results and unresolved work after release | Follow-up decisions and any exceptions |
Approval does not need to mean asking permission before every edit.
An agent should complete the work already authorized and present a concrete result. It should stop when a new decision changes the contract, crosses an authorization boundary or requires evidence it cannot obtain.
Written rules also need suitable repository permissions, review and operational discipline. A Markdown instruction by itself is not a technical security boundary.
Turn assumptions into testable decisions¶
Authentication and interactive behavior illustrate why architecture review matters. For example, an operation that prompts for credentials may work in an interactive terminal but hang or fail in unattended CI. The execution context belongs in the contract and its tests.
I aim to make those assumptions explicit before implementation and verify them against supported environments. A reviewed proposal is still only a design decision; it is not proof that the released behavior works.
Qualification is environment-specific¶
A successful test in one environment cannot prove behavior across other operating systems, credential mechanisms or live provider integrations. Claims should identify what was actually tested, which artifacts were used and what remains unverified.
AI-assisted changes require the same discipline as other engineering changes: reviewable diffs, bounded permissions and evidence matched to the claimed outcome.
Close the loop with release evidence¶
An implementation being merged does not by itself establish that the relevant user problem is resolved. I distinguish code changes from release qualification and use post-release findings to inform future work.
Evidence should make the result independently inspectable, including limitations and unresolved cases. The goal is a trustworthy outcome, not a polished automated report.
Context must survive a session ending¶
AI sessions end and development context changes. Durable decisions, review history and qualification records let work resume without relying on an earlier conversation.
Those records do not grant authority or fill in missing decisions. The companion article GitHub as Agent Context explains the broader principle of keeping engineering context alongside versioned work.
A small starting point for another solo founder¶
You do not need to copy the whole repository layout.
Start with a decision log, an issue tracker, a backlog folder and one release proposal.
For each release, write:
- The user outcome and primary contract.
- What is included and deliberately excluded.
- The source baseline and dependencies.
- The acceptance tests and required environments.
- Who may approve, implement, merge and publish.
- What evidence must remain, including issue disposition after release.
Use research or review automation only when you have capacity to inspect its output.
The cost of this system is real: stale documents, duplicate proposals, excessive process and noisy findings can consume the time agents were supposed to save. Keep one canonical decision, update existing work, and add a rule when a concrete failure justifies it.
The measure of success remains whether a released change delivers a verifiable engineering outcome.
The useful outcome is a development process that preserves human judgment and leaves enough evidence for someone else to continue.