Reviewers can challenge code and plans, but suggestions do not silently become accepted repository state.
Case study · Agentic engineering
Agents produce drafts.
Evidence earns trust.
I use coding agents as bounded participants in a delivery system: one owner controls repository state, independent reviewers challenge the work, tests and browser checks produce proof, and consequential actions stop for human approval.
System boundary
The model is not the trust boundary.
One repo ownerMutable state stays under a single accountable workflow
Bounded reviewSecond models inspect scoped packets without repository authority
Explicit risk gatesData, money, secrets, sends, and deploys stop for approval
Completion is attached to inspectable evidence—a focused test, reviewed diff, browser capture, health check, or release record—not a confident agent status message.
Concrete workflow evolution
From manual coordination to controlled delivery.
01
EarlierI carried assignments and state between separate model conversations.
CurrentOne owner holds mutable state; reviewers receive exact file-backed questions.
02
EarlierCompletion leaned too heavily on an agent's status summary.
CurrentTests, diffs, screenshots, and release records accompany completion claims.
03
EarlierRisk boundaries were implicit in conversation.
CurrentConsequential actions route to explicit, visible approval gates.
Hard engineering decisions
Control beats simulated consensus.
Verification artifacts falsify overconfident summaries and make handoffs inspectable.
Routine work can continue autonomously while unclear financial, data, credential, send, or deployment impact stops.
Origin evidence · around mid-2024
Agent Beach made coordination visible.
The original working local interface was called Agent Beach. It originated around mid-2024 and visualized role assignments, Strategy/Build/QA/Launch zones, progress, handoffs, and artifacts.
The public-safe presentation was completed May 16, 2026. It uses fictional local state and does not call model-provider APIs. Later model labels are illustrative and are not timestamp evidence for the original build.
Inspect Agent Beach ↗
Public/private boundary
Inspectable patterns without leaking operational context.
- Public: workflow architecture, risk tiers, ownership rules, fictional examples, and source repositories.
- Public: Agent Beach screenshots and browser-tested simulated state.
- Private: production repositories, customer data, credentials, raw prompts, transcripts, and runbooks.
- Explicit limit: Agent Beach is historical origin evidence, not the current runtime or a live multi-provider orchestrator.