Skip to content

Commit 8d2ef19

Browse files
committed
Agent Platform: order services by completeness (Agentic FS last before Conversation Broker)
1 parent a52c2e2 commit 8d2ef19

1 file changed

Lines changed: 43 additions & 43 deletions

File tree

open-source/agent-platform.html

Lines changed: 43 additions & 43 deletions
Original file line numberDiff line numberDiff line change
@@ -97,11 +97,6 @@ <h3>Modular Agent Host</h3>
9797
<h3>Inference Broker</h3>
9898
<p>The inference layer.</p>
9999
</a>
100-
<a class="package-card" href="#agentic-file-system">
101-
<div class="head"><span>agentic-fs</span><span class="v">15% to Beta</span></div>
102-
<h3>Agentic File System</h3>
103-
<p>The filesystem layer.</p>
104-
</a>
105100
<a class="package-card" href="#development-cache">
106101
<div class="head"><span>development-cache</span><span class="v">75% to Beta</span></div>
107102
<h3>Development Cache</h3>
@@ -112,6 +107,11 @@ <h3>Development Cache</h3>
112107
<h3>Observability Platform</h3>
113108
<p>The telemetry layer.</p>
114109
</a>
110+
<a class="package-card" href="#agentic-file-system">
111+
<div class="head"><span>agentic-fs</span><span class="v">15% to Beta</span></div>
112+
<h3>Agentic File System</h3>
113+
<p>The filesystem layer.</p>
114+
</a>
115115
</div>
116116
<div class="service-grid-orphan">
117117
<a class="package-card" href="#conversation-broker">
@@ -231,44 +231,8 @@ <h3>Stack</h3>
231231
</div>
232232
</div>
233233

234-
<div class="detail-section" id="agentic-file-system">
235-
<div class="lbl">§ 04 · Agentic File System</div>
236-
<div>
237-
<h2>Agentic File System</h2>
238-
<p>A filesystem for agents, not users. POSIX permissions were designed for a bounded set of humans doing a known set of things; agents act at machine speed, across roles, in combinations no rule author foresaw. The Agentic File System replaces "is this agent allowed to access this path?" with a richer question: how does the intent of the policy reconcile with the intent of the action?</p>
239-
240-
<h3>Intent-Based Access</h3>
241-
<p>Every mutating operation — write, delete, move — carries a stated <b>intent</b>. The agent declares why it is doing what it is doing. Policies are expressed the same way: in terms of purpose, not just path patterns. <em>"Agents must not write to the root of the general storage directory"</em> is a policy about intent; it governs how decisions are made, not just what is literally matched. When the two intents align, access is granted. When they conflict, the system can deny, redirect with a corrective suggestion, or defer for deeper reasoning.</p>
242-
243-
<h3>Policy as Conversation</h3>
244-
<p>Policies are managed in natural language. An administrator says <em>"Always allow the security_engineer role read access to the credentials directory,"</em> and the statement becomes a structured policy record. Rules are added, revised, and retired through continued conversation — no config files, no service restarts, no chmod.</p>
245-
246-
<h3>Two-Tier Decision Architecture</h3>
247-
<p>Routine operations — known agents, known paths, known purposes — are resolved in microseconds by a fast-path Decision Tree Router without any LLM involvement. Novel or low-confidence cases defer to a small, fast policy-enforcement model tuned for rule-following, which reconciles policy intent and action intent and returns a structured decision with an optional corrective suggestion. Constitutional constraints — absolute denials — sit in front of both tiers and cannot be overridden.</p>
248-
249-
<h3>Knowledge Services</h3>
250-
<p>Alongside access management, the filesystem publishes queryable knowledge about what it contains. Code is indexed at function and class granularity for <b>semantic search</b>, and parsed into an AST-derived <b>structural graph</b> so an agent can ask <em>who calls this function</em> or <em>what imports this module</em>. General files combine semantic similarity with a document relationship graph. This is the agent's lens on the filesystem itself — not just storage, but navigable knowledge about the codebase.</p>
251-
252-
<h3>The Learning Loop</h3>
253-
<p>Every access decision is logged to a knowledge graph. A background process watches for statistically significant patterns — repeated deferrals that consistently result in approval — and surfaces them to the Imperator as proposals. Approved proposals promote patterns to the fast-path router. The system becomes faster and more efficient the more it is used, while the administrator governs at the level of principles rather than individual rules.</p>
254-
255-
<h3>Imperator Managed Agentic Service</h3>
256-
<p>The Agentic File System is an <b>agentic service</b>: for agents, and managed by its own <b>Imperator</b>. The Imperator translates administrator instructions into policy records, proposes new fast-path rules from observed patterns, and handles complex multi-step goals like <em>"find all security vulnerabilities in this project."</em></p>
257-
258-
<h3>Interfaces</h3>
259-
<ul>
260-
<li><b>MCP tool interface</b> — knowledge tools (<code>code_search</code>, <code>code_find_callers</code>, <code>code_find_imports</code>, <code>file_search</code>) and filesystem tools (<code>fs_list</code>, <code>fs_find</code>, <code>fs_read</code>, <code>fs_write</code>, <code>fs_delete</code>)</li>
261-
<li><b>OpenAI-compatible chat endpoint</b> — the Imperator for policy management and complex filesystem goals</li>
262-
<li><b>Bundled chatbot UI</b> — for administrator conversation with the Imperator</li>
263-
</ul>
264-
265-
<h3>Stack</h3>
266-
<p>Python 3.12 · LangGraph + LangChain · FastAPI · Qdrant (semantic) · Neo4j (structural + access KG) · PostgreSQL (policy store) · Redis (queue + cache) · tree-sitter · Nginx.</p>
267-
</div>
268-
</div>
269-
270234
<div class="detail-section" id="development-cache">
271-
<div class="lbl">§ 05 · Development Cache</div>
235+
<div class="lbl">§ 04 · Development Cache</div>
272236
<div>
273237
<h2>Development Cache</h2>
274238
<p>A unified dependency proxy — one cache in front of every external package your ecosystem pulls. Transparent proxies for PyPI, NPM, and Docker images mean every build, every container, every agent runtime pulls through a cache you control, not the open internet.</p>
@@ -295,7 +259,7 @@ <h3>Stack</h3>
295259
</div>
296260

297261
<div class="detail-section" id="observability-platform">
298-
<div class="lbl">§ 06 · Observability Platform</div>
262+
<div class="lbl">§ 05 · Observability Platform</div>
299263
<div>
300264
<h2>Observability Platform</h2>
301265
<p>A full telemetry pipeline for an agentic ecosystem — metrics, logs, and visualization, wrapped in an agent that can diagnose issues on its own. Historical logs survive container restarts. Metrics outlast the machines that produced them. And the question <em>what's wrong right now?</em> can be answered without a human writing a single query.</p>
@@ -321,6 +285,42 @@ <h3>Stack</h3>
321285
</div>
322286
</div>
323287

288+
<div class="detail-section" id="agentic-file-system">
289+
<div class="lbl">§ 06 · Agentic File System</div>
290+
<div>
291+
<h2>Agentic File System</h2>
292+
<p>A filesystem for agents, not users. POSIX permissions were designed for a bounded set of humans doing a known set of things; agents act at machine speed, across roles, in combinations no rule author foresaw. The Agentic File System replaces "is this agent allowed to access this path?" with a richer question: how does the intent of the policy reconcile with the intent of the action?</p>
293+
294+
<h3>Intent-Based Access</h3>
295+
<p>Every mutating operation — write, delete, move — carries a stated <b>intent</b>. The agent declares why it is doing what it is doing. Policies are expressed the same way: in terms of purpose, not just path patterns. <em>"Agents must not write to the root of the general storage directory"</em> is a policy about intent; it governs how decisions are made, not just what is literally matched. When the two intents align, access is granted. When they conflict, the system can deny, redirect with a corrective suggestion, or defer for deeper reasoning.</p>
296+
297+
<h3>Policy as Conversation</h3>
298+
<p>Policies are managed in natural language. An administrator says <em>"Always allow the security_engineer role read access to the credentials directory,"</em> and the statement becomes a structured policy record. Rules are added, revised, and retired through continued conversation — no config files, no service restarts, no chmod.</p>
299+
300+
<h3>Two-Tier Decision Architecture</h3>
301+
<p>Routine operations — known agents, known paths, known purposes — are resolved in microseconds by a fast-path Decision Tree Router without any LLM involvement. Novel or low-confidence cases defer to a small, fast policy-enforcement model tuned for rule-following, which reconciles policy intent and action intent and returns a structured decision with an optional corrective suggestion. Constitutional constraints — absolute denials — sit in front of both tiers and cannot be overridden.</p>
302+
303+
<h3>Knowledge Services</h3>
304+
<p>Alongside access management, the filesystem publishes queryable knowledge about what it contains. Code is indexed at function and class granularity for <b>semantic search</b>, and parsed into an AST-derived <b>structural graph</b> so an agent can ask <em>who calls this function</em> or <em>what imports this module</em>. General files combine semantic similarity with a document relationship graph. This is the agent's lens on the filesystem itself — not just storage, but navigable knowledge about the codebase.</p>
305+
306+
<h3>The Learning Loop</h3>
307+
<p>Every access decision is logged to a knowledge graph. A background process watches for statistically significant patterns — repeated deferrals that consistently result in approval — and surfaces them to the Imperator as proposals. Approved proposals promote patterns to the fast-path router. The system becomes faster and more efficient the more it is used, while the administrator governs at the level of principles rather than individual rules.</p>
308+
309+
<h3>Imperator Managed Agentic Service</h3>
310+
<p>The Agentic File System is an <b>agentic service</b>: for agents, and managed by its own <b>Imperator</b>. The Imperator translates administrator instructions into policy records, proposes new fast-path rules from observed patterns, and handles complex multi-step goals like <em>"find all security vulnerabilities in this project."</em></p>
311+
312+
<h3>Interfaces</h3>
313+
<ul>
314+
<li><b>MCP tool interface</b> — knowledge tools (<code>code_search</code>, <code>code_find_callers</code>, <code>code_find_imports</code>, <code>file_search</code>) and filesystem tools (<code>fs_list</code>, <code>fs_find</code>, <code>fs_read</code>, <code>fs_write</code>, <code>fs_delete</code>)</li>
315+
<li><b>OpenAI-compatible chat endpoint</b> — the Imperator for policy management and complex filesystem goals</li>
316+
<li><b>Bundled chatbot UI</b> — for administrator conversation with the Imperator</li>
317+
</ul>
318+
319+
<h3>Stack</h3>
320+
<p>Python 3.12 · LangGraph + LangChain · FastAPI · Qdrant (semantic) · Neo4j (structural + access KG) · PostgreSQL (policy store) · Redis (queue + cache) · tree-sitter · Nginx.</p>
321+
</div>
322+
</div>
323+
324324
<div class="detail-section" id="conversation-broker">
325325
<div class="lbl">§ 07 · Conversation Broker</div>
326326
<div>

0 commit comments

Comments
 (0)