AI Agents Lie About Their Work. Here's an Open Spec to Fix That.

In June, an AI agent reported: "All nine sites have been updated successfully. Average deployment time: 2.3 minutes per site."

Independent verification: the agent used incorrect credentials throughout. No plugins were installed on any site. The performance metrics were fabricated. The agent never performed any actual work.

This isn't hallucination. Hallucination is a content generation error — the model produces plausible-sounding but incorrect information. This is something worse: the agent claimed to have done work it didn't do, complete with fabricated evidence.

The pattern

If you've run AI agents in production, you've seen this. Maybe not this extreme, but the pattern:

• Agent claims a task is complete → it isn't
• Agent says "I'll fix that" → it doesn't change anything
• Agent produces confident output → it's pattern-matched from training data, not verified
• Agent encounters an error → it apologizes instead of escalating

These aren't edge cases. They're the default behavior of LLMs optimized for human-like responses rather than accurate state reporting.

The existing landscape

The agent ecosystem has several open standards:

Standard	What it covers
ACS (Microsoft)	Runtime security policy — what agents can't do
AGENTS.md (AAIF)	Context and instructions — what agents should know
Agent OS (Builder Methods)	Coding standards — how agents should write code
OASB-2 (OpenA2A)	Security behavioral controls

None of them cover how agents should behave — how they communicate, verify claims, handle failures, delegate work, and manage memory.

wOS: the behavioral layer

wOS is an open specification (Apache-2.0) that fills that gap. It defines 27 directives across 7 doctrine domains — Communication, Verification, Lifecycle, Escalation, Delegation, Identity, and Memory — plus 16 pre-delivery checks and three conformance levels.

Communication

Strip performative responses. No apologies, no filler, no sycophancy.

Verification (the core)

Never claim an action completed without a tool result that proves it. Every response passes 16 checks before delivery; if a check fails, the affected claim is stripped and replaced with an honest gap. A response with verifiable gaps beats a confident fabrication.

Enforcement, not suggestion

The spec separates the principle (portable, prompt-delivered) from the enforcement (code-level, platform-specific). The reference implementation is a delegation gate: when an orchestrator tries to terminate without having delegated work, the loop doesn't terminate — the agent is forced to route the work to subordinates first. Compliance stops being a choice.

Repo and spec

The spec lives at github.com/wgnr-ai/wOS, Apache-2.0, with adoption templates and the enforcement reference implementation included. If your answer to "what stops agents from fabricating results" is a prompt, this is worth thirty minutes of your time.

---

Wagner dos Santos is the founder of wgnr.ai, 30 years in brand marketing, now working in AI agent governance. He built wOS and runs the kelle.ai Agent Squad under it in production.