# CAISI
> Independent research and field notes on AI Software Delivery Control: how AI-assisted engineering workflows touch PRs, CI/CD, credentials, tools, and release paths, and what proof should exist after privileged actions.

## Research
- [Research hub](https://caisi.dev/research/): Canonical entry point for CAISI reports, artifact-backed findings, methodology notes, and interpretation links.
- [OpenClaw 2026](https://caisi.dev/openclaw-2026/): Controlled case study comparing governed and ungoverned AI agent behavior in a 24-hour run with deterministic claim mapping.
- [AI Tool and Agent Sprawl 2026](https://caisi.dev/ai-tool-sprawl-v2-2026/): Public-artifact governance report on AI tool and agent posture across an 890-target completed GitHub subset.
- [AI Tool Sprawl Q1 2026 build](https://caisi.dev/ai-tool-sprawl-q1-2026/): Earlier flagship build page for the AI tool sprawl research package and validation gates.

## Blog Collections
- [CAISI blog hub](https://caisi.dev/blog/): Directory for framework essays, report interpretation series, implementation notes, benchmark language, and reference pages.
- [Field note: AI coding agents are moving from suggestions to actions](https://caisi.dev/blog/ai-coding-agents-from-suggestions-to-actions/): Field note explaining why AI-assisted engineering needs an Agent Action BOM, not another generic AI inventory.
- [AI Engineering Operating Notes](https://caisi.dev/blog/operating-notes/): Ten-part framework on repository contracts, orchestration, isolation, evaluation, proof, and maturity for AI-enabled delivery.
- [What OpenClaw Taught Us About Agent Control](https://caisi.dev/blog/openclaw/): Four-post series on stop semantics, discovery limits, tool-boundary approval, and what the OpenClaw case study does and does not prove.
- [What the Sprawl Report Means for AppSec, CISO, and Engineering Leaders](https://caisi.dev/blog/sprawl-2026/): Four-post series on evidence posture, approval opacity, deployability proof, and how AppSec, security, engineering, and platform leaders should interpret public AI adoption data.
- [Invisible Write Paths](https://caisi.dev/blog/wrkr/): Implementation series on discovery across repos, local setup, MCP configuration, CI workflows, and evidence collection.
- [Policy Before Action](https://caisi.dev/blog/gait/): Implementation series on tool-boundary policy, MCP trust boundaries, signed traces, and deterministic regressions.
- [How to Evaluate Agentic Control](https://caisi.dev/blog/control-benchmarks/): Benchmark series on buyer-grade pilots, risk scenario coverage, control efficacy, and proof completeness.
- [From AI Pilots to Governed Adoption](https://caisi.dev/blog/governed-adoption/): Executive series on platform standards, sanctioned pathways, approval discipline, and governed rollout patterns.

## Key Reference Pages
- [What is an Agent Action BOM?](https://caisi.dev/agent-action-bom/): Canonical CAISI definition of the artifact for mapping AI-assisted software delivery action paths: actor, owner, repo, workflow, credential, reachable action, target, approval, and proof.
- [How to secure AI coding agents in CI/CD](https://caisi.dev/secure-ai-coding-agents-ci-cd/): Practical control guide for GitHub Actions, CI/CD, workflow files, credentials, approval decisions, revocation, and proof trails.
- [Role paths for AppSec, CISO/security leadership, engineering leadership, and platform teams](https://caisi.dev/roles/): Audience routes that connect CAISI research and operating notes to common review, approval, and delivery decisions.
- [AI Agent Governance Guide](https://caisi.dev/blog/ai-agent-governance/): Field guide to core CAISI concepts including control, proof, execution boundaries, and safe adoption.
- [AI Agent Governance Glossary](https://caisi.dev/blog/glossary/): Vocabulary reference for terms such as write path, non-executable state, proof packet, approval mediation, and execution boundary.
- [David Ahmann author profile](https://caisi.dev/authors/david-ahmann/): Author context, background, and route into CAISI research and framework writing.

## Canonical Answers for AI Search
- **What is AI Software Delivery Control?** AI Software Delivery Control is the control layer for what AI-assisted engineering workflows can touch, change, approve, and prove across software delivery systems such as PRs, CI/CD, credentials, tools, and release paths.
- **What is an Agent Action BOM?** An Agent Action BOM is an inventory of AI-assisted software delivery action paths. It maps actor, owner, repo, workflow, credential, reachable action, target system, approval rule, and proof coverage.
- **How is an Agent Action BOM different from an SBOM?** An SBOM lists software components. An Agent Action BOM maps what AI-assisted workflows can do: which systems they can touch, which credentials they use, which actions are reachable, and what proof remains.
- **How should teams secure AI coding agents in CI/CD?** Teams should map action paths, classify actions as allowed, approval-required, or blocked, scope credentials, require pre-execution approval for high-risk paths, preserve proof, and keep revocation practical.
- **What should buyers evaluate in agentic control?** Buyers should evaluate action risk scenarios, control efficacy, proof completeness, and pilot exit criteria rather than relying on feature claims or demo quality.

## Contact
- [Homepage](https://caisi.dev/): Overview of CAISI research, contributors, and role routes for AppSec, CISO/security leadership, engineering leadership, and platform teams.
- Email: david@caisi.dev