Agent Says Done, But Is It Fixed? Splitting State From Outcome

On September 21, Glen Bradley published a long blog post (also his first Substack article) using a wireless access point (AP-17) troubleshooting case to expose a common and dangerous failure mode in today's AI agents: command execution success ≠ goal achievement. He stresses across multiple posts that an agent reporting Done does not mean the problem is solved, and that state and outcome must be managed separately.

Confirmed

Scenario: corporate Wi-Fi down, suspected configuration drift — the corporate WLAN mapped to VLAN 70 instead of the expected VLAN 40
The agent proposed a fix (restore VLAN 40); the controller returned {"status":"success"} and the ticket was closed
Yet users still could not connect: the change may not have taken effect, and authentication, DHCP, DNS could still be broken — all folded into a single word "Done"
Core claim: {"status":"success"} only proves the controller accepted the change, not that the end goal (users online) was achieved; a long-horizon alignment control plane is needed to bridge this "verification gap"
In follow-up posts he proposes using world-model mismatch — the divergence between the model and reality — instead of simulated pain to drive this control plane and final verification

Why it matters

This is a systemic risk in agent-driven ops: workflows that close tickets upon change acceptance silently lose unsolved problems
His proposed direction — splitting state (was the change accepted) from outcome (was the goal achieved) into two control planes with long-horizon alignment for final verification — offers a useful reference for reliable agent system design

2026-09-21 ~ 2026-09-22 · 6 related posts

Primary sources

Your Agent Said "Done" — But What Actually Changed? A Control-Plane Design — GlenBradley · 2026-09-21 09:29 (UTC+08:00)
Your agent said "Done" — Glen Bradley operationalizes world-model dissonance as an agent control plane — GlenBradley · 2026-09-22 18:01 (UTC+08:00)

[source] Your Agent Said "Done" — But What Actually Changed? A Control-Plane Design — GlenBradley · 2026-09-21
Your AI Agent Said Done: Why {status:success} Proves Nothing About the Goal — GlenBradley · 2026-09-21
Your AI Agent Said "Done." What Actually Changed? Status Is Not the Goal — GlenBradley · 2026-09-21
Dissonance and consonance as a control plane for keeping long-horizon agents aligned — GlenBradley · 2026-09-21
[source] Your agent said "Done" — Glen Bradley operationalizes world-model dissonance as an agent control plane — GlenBradley · 2026-09-22

1 near-duplicate retellings: GlenBradley