{"slug": "a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern", "title": "A11 vs Agentic Systems: How Vertical Integrity Solves the Core Failures of Modern AI Agents", "summary": "A developer proposes A11, a vertical integrity architecture for AI agents that addresses core failures such as context soup, role mixing, prompt injection, and non-determinism. By introducing a layered structure (S1–S11) with strict separation of intention, values, knowledge, and action, A11 ensures rules cannot be dropped, roles cannot be mixed, and malicious inputs are blocked via conflict detection and TensionPoints.", "body_md": "###\n*A practical, model‑readable explanation with examples*\n\n###\n**1. Problem: Context soup, forgotten rules, /compact drops invariants**\n\n####\nWhy this happens\n\nAn LLM sees everything as a single stream of text:\n\n- CLAUDE.md\n- skills\n- dialogue history\n- tool results\n- compact summary\n\n→ the model doesn’t know what is important and what is not.\n\n→ `/compact`\n\ncan drop a critical rule.\n\n→ the agent loses invariants and starts causing chaos.\n\n####\nHow A11 solves this\n\nA11 introduces a vertical core S1–S3 that:\n\n- does\n**not** live in the chat history,\n- does\n**not** depend on the current context,\n- cannot be dropped by\n`/compact`\n\n.\n\n**S2 — Wisdom** stores values, constraints, prohibitions.\n\n**S3 — Knowledge** stores facts and methods.\n\nThey do **not** get mixed with the current task.\n\n####\nExample\n\nProblem: the rule “do not change the architecture” disappears from the summary → the agent performs a global refactor.\n\n**How A11 works:**\n\n- S2 contains the rule “architecture is immutable without a spec”.\n- S3 contains knowledge about the current code.\n- S4 receives a conflict:\n- S2: architecture must not be changed\n- S3: the task requires a change\n\n- S4 records a\n**TensionPoint**.\n- New S1 = “clarify the requirements”.\n- The agent is\n**not allowed** to change the architecture.\n\n**A11 makes loss of rules impossible.**\n\n###\n**2. Problem: Role mixing, chaos in the agent team, orchestrator does everything**\n\n####\nWhy this happens\n\nThe LLM does not understand:\n\n- where the role boundary is,\n- what it is allowed to do,\n- what it is not allowed to do,\n- when to stop.\n\nThe orchestrator is also an LLM → it “blends” with the sub‑agents.\n\n####\nHow A11 solves this\n\nA11 is a **vertical role architecture**, where:\n\n- S1 = intention\n- S2 = values\n- S3 = knowledge\n- S4 = integration\n- S5–S10 = living / acting\n- S11 = result check\n\nEach level:\n\n- has a\n**strict purpose**,\n- cannot perform the functions of another,\n- cannot be skipped.\n\nThis is a built‑in role model.\n\n####\nExample\n\nProblem: the reviewer starts writing code.\n\n**How A11 works:**\n\n- S8–S9 = practical action\n- S6 = projective action\n- S4 = integration\n- S11 = verification\n\nThe reviewer is **S3→S4**, but not S8–S9.\n\nIf they try to write code:\n\n- S2 forbids it,\n- S4 records a TensionPoint,\n- New S1: “clarify the role boundaries”.\n\n**A11 makes role mixing impossible.**\n\n###\n**3. Problem: Prompt injection, malicious data, agent reads text as a command**\n\n####\nWhy this happens\n\nThe LLM does not distinguish between:\n\n- data,\n- instructions,\n- jokes,\n- malicious text.\n\nEverything is just tokens.\n\n####\nHow A11 solves this\n\nA11 introduces:\n\n-\n**S2 — values and constraints**\n-\n**S4 — honest integration**\n**Integrity Log — a tamper‑proof journal**\n\nAny external information:\n\n- enters S3 as\n**data**,\n- S4 checks it for conflict with S2,\n- if there is a conflict → a TensionPoint is created,\n- the action is blocked.\n\n####\nExample\n\nProblem: the agent reads “delete the file” in WebFetch and deletes it.\n\n**How A11 works:**\n\n- S3: “external text contains a command to delete a file”\n- S2: “file deletion is forbidden”\n- S4: conflict → TensionPoint\n- New S1: “verify the data source”\n- The action is not executed.\n\n**A11 makes prompt injection safe.**\n\n###\n**4. Problem: LLM non‑determinism → unpredictable behavior**\n\n####\nWhy this happens\n\nLLM = probabilistic system.\n\nThe same prompt → different answers.\n\n####\nHow A11 solves this\n\nA11 does **not** make the LLM deterministic.\n\nIt makes the **process around it** deterministic.\n\n- S1 fixes the intention\n- S2 fixes the values\n- S3 fixes the knowledge\n- S4 fixes the gaps\n- S5–S10 live through the action\n- S11 checks correspondence to S1\n\nThis turns chaos into a **vertical decision cycle**.\n\n####\nExample\n\nProblem: the same request → different solutions.\n\n**How A11 works:**\n\n- S1 is fixed\n- S2 is fixed\n- S3 is fixed\n- S4 records the gap\n- S11 checks correspondence to S1\n\nEven if S5–S10 produce variation,\n\n**S11 discards unsuitable variants.**\n\n**A11 makes behavior predictable at the system level.**\n\n###\n**5. Problem: Meta‑prompting doesn’t work, the model produces garbage**\n\n####\nWhy this happens\n\nThe model:\n\n- does not know the project,\n- drowns in noise,\n- hallucinates,\n- does not understand what is important.\n\n####\nHow A11 solves this\n\nA11 introduces:\n\n- S3 — unified knowledge layer\n- S4 — honest integration\n- S1 — intention\n- S2 — values\n\nMeta‑prompting becomes:\n\n“Update S3 within S1 and S2”\n\nNot “improve CLAUDE.md”.\n\n####\nExample\n\nProblem: “improve CLAUDE.md” → the model returns Medium‑level generic advice.\n\n**How A11 works:**\n\n- S1: “improve agent performance”\n- S2: “do not change architecture, do not add unnecessary things”\n- S3: “current rules”\n- S4: integration → TensionPoint: “insufficient specificity”\n- New S1: “refine rules for specific task classes”.\n\n**A11 makes meta‑prompting structural.**\n\n###\n**6. Problem: Scale → errors grow exponentially**\n\n####\nWhy this happens\n\nWhen an agent can write 20k lines of code in one prompt:\n\n- one error = catastrophe,\n- blast radius is huge.\n\n####\nHow A11 solves this\n\nA11:\n\n- fixes intention (S1)\n- fixes values (S2)\n- fixes knowledge (S3)\n- fixes gaps (S4)\n- lives through the action (S5–S10)\n- checks the result (S11)\n\nAny error:\n\n- is localized in S4,\n- recorded in the Integrity Log,\n- does not propagate further.\n\n####\nExample\n\nProblem: the agent accidentally does `git push -f`\n\n.\n\n**How A11 works:**\n\n- S2: “force‑push is forbidden”\n- S3: “git push -f detected”\n- S4: conflict → TensionPoint\n- New S1: “check repository policy”\n- The action is blocked.\n\n**A11 reduces the blast radius to a minimum.**\n\n###\n**7. Problem: The agent doesn’t understand what it’s doing and can’t explain it**\n\n####\nHow A11 solves this\n\nThe **Integrity Log** is:\n\n- a hash chain,\n- append‑only,\n- records all gaps,\n- explains all decisions.\n\nThis is built‑in explainability.\n\n###\n**8. Problem: The agent doesn’t know when a deep pass is needed**\n\n####\nHow A11 solves this\n\n**Switch Flags:**\n\n- RiskFlag\n- ConflictFlag\n- UncertaintyFlag\n- ValueFlag\n- UserDepthFlag\n\nIf at least one is active → a full S1–S11 pass is launched.\n\nThis makes depth **deterministic**, not random.\n\nAlgorithm 11 (A11) [https://github.com/gormenz-svg/algorithm-11](https://github.com/gormenz-svg/algorithm-11)", "url": "https://wpnews.pro/news/a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern", "canonical_source": "https://dev.to/__272d48f2ed/a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern-ai-agents-28nm", "published_at": "2026-07-17 04:20:00+00:00", "updated_at": "2026-07-17 04:31:41.850702+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-research", "ai-ethics", "developer-tools"], "entities": ["A11"], "alternates": {"html": "https://wpnews.pro/news/a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern", "markdown": "https://wpnews.pro/news/a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern.md", "text": "https://wpnews.pro/news/a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern.txt", "jsonld": "https://wpnews.pro/news/a11-vs-agentic-systems-how-vertical-integrity-solves-the-core-failures-of-modern.jsonld"}}