State of Agent Readiness - May 2026 | Production AI Institute

Repositories20

Average coverage38%

Human oversight20%

Observability55%

Edition finding

Deployment controls were visible; autonomy governance was not yet routine.

In this edition, the strongest visible domain was D5 Deployment control (85% of scanned repositories showed at least one signal). The weakest visible domain was D6 Human oversight (20%). This means maintainers can often show build and release discipline, but many still need explicit approval gates, eval evidence, and operational incident procedures.

Domain coverage

D165%

D240%

D330%

D455%

D585%

D620%

D770%

D860%

What changed into action

Publish a docs/production-ai-readiness.md evidence map that links each PSF domain to the relevant repository artifact.
Add an evals folder with regression cases, scoring thresholds, and the current model or prompt version under test.
Document human approval gates for high-impact actions, including who approves, what is logged, and what fails closed.
Add an incident response note, degraded-mode path, and provider fallback procedure.

Findings

Deployment control evidence appeared in 17 of 20 repositories, most often through CI workflows, release automation, or versioned deployment paths.
Human oversight evidence appeared in 4 of 20 repositories, making autonomy gates the clearest visible gap in this edition.
No repository in this public sample exposed eval evidence that matched the current scanner signal pattern.
Security and provider resilience signals were more visible than data stewardship, suggesting maintainers often document engineering controls before data governance controls.

Top visible evidence examples

Evidence pack builder

serac-labs/seracAI observability instrumentation, Human approval gates, Security policy and secret hygiene

84%

vm0-ai/vm0AI observability instrumentation, Security policy and secret hygiene, Provider fallback or degraded mode

77%

Icarus603/claude-codeAI observability instrumentation, Human approval gates, Security policy and secret hygiene

76%

HankHuang0516/EClawAI observability instrumentation, Human approval gates, Security policy and secret hygiene

63%

holaboss-ai/holaOSAI observability instrumentation, Security policy and secret hygiene, Provider fallback or degraded mode

59%

bug-ops/zephAI observability instrumentation, Security policy and secret hygiene, Provider fallback or degraded mode

58%

slicenferqin/xuanpuAI observability instrumentation, Security policy and secret hygiene, Provider fallback or degraded mode

54%

hrygo/hotplexAI observability instrumentation, Human approval gates, Security policy and secret hygiene

53%

Source and method

Repository discovery used GitHub public repository search. The scanner reviewed public repository metadata and public file-tree paths only. It did not clone repositories, inspect private code, or certify the projects listed here.

topic:ai-agent archived:false fork:false stars:>=5
topic:agentic-ai archived:false fork:false stars:>=5
topic:llm-agent archived:false fork:false stars:>=5
topic:mcp-server archived:false fork:false stars:>=5
"ai agent" in:name,description,readme archived:false fork:false stars:>=5

Citation note

Production AI Institute. "State of Agent Readiness - May 2026." Frozen May 13, 2026. Available at https://www.productionai.institute/agent-readiness/benchmark/2026-05.

This edition is an immutable archival snapshot. It does not change when the live benchmark changes.
The scan uses public repository metadata and public file-tree paths. It does not clone private code and it is not a certification or endorsement.
Authenticated production runs should use a server-side GitHub token to raise API limits before freezing future monthly editions.