Authored: 2026-05-13 | Agent: SCOUT | Scope: April-May 2026 Claude Code delta vs TITAN current posture
---
Anthropic's hosted filesystem-backed memory layer for agents: entries survive across sessions, support read-write scopes, per-write audit logs, rollback, and parallel-safe concurrency. TITAN currently hand-rolls file-based memory in F:/TITAN/knowledge/. The Managed Memory API lets agents mount a versioned memory store with export via Claude Console. Early adopters report 97% fewer first-pass errors and 27% cost reduction (Rakuten pilot). Integration effort: medium — needs Managed Agents API access + migration of hot-memory writes from the current file approach to API calls. Still beta; suitable for non-production pilots immediately.
type: "mcp_tool" Hook Output (v2.1.118)Hooks can now invoke MCP tools directly as their output without spawning a subprocess. TITAN's PostToolUse hook currently writes JSON files and shells out to Python scripts — the older pattern. Switching hook outputs to type: "mcp_tool" eliminates subprocess overhead and lets hooks participate natively in the MCP call graph. Integration effort: low — update existing hook JSON responses in F:/TITAN/hooks/.
The PreCompact hook now supports {"decision": "block"} or exit code 2 to abort compaction entirely. TITAN has no PreCompact hook installed. Adding one that (a) saves hot context to a staging file before compaction and (b) blocks compaction in agent sessions where state continuity is critical (symphony generation, newsletter send) would eliminate the "forgetful after compact" failure mode already documented in TITAN's internals memo. Integration effort: low — single hook file + snapshot logic.
The Monitor tool streams every stdout line from a background process back to Claude as notifications — no polling loop needed. TITAN cron tasks and symphony jobs use Bash with run_in_background and periodic re-checks. Replacing those patterns with Monitor gives real-time task observability and lets agents react to failures mid-flight rather than discovering them on the next check-in. Integration effort: low — swap Bash background calls with Monitor in skills that spawn long-running processes.
Claude Code now ships a first-class Jupyter notebook editing tool. TITAN has no notebook layer. Given Harnoor's data engineering stack (DuckDB, Polars, Trino, Iceberg), a notebook-based exploratory analysis skill (e.g., /explore that scaffolds a DuckDB notebook from a schema) would be high-value and requires no extra plumbing since the tool is already built in. Integration effort: medium — new skill + CLAUDE.md integration only.
sampling/createMessage)MCP servers can now request completions back from the Claude client — enabling MCP sub-agents to loop Claude's reasoning into their own processing. TITAN's MCP integrations are currently one-way: Claude calls them. Bidirectional sampling unlocks patterns like a doc-indexing MCP server that asks Claude to interpret ambiguous results before returning them, and reduces context-dump token cost by ~30% vs. pasting raw docs into context. Integration effort: high — requires MCP server changes + new protocol handling.
The formula ET = (m x I) + (0.5 x m x C) + (4 x m x O) (model-weighted, output-heavy) gives a single cost-equivalent unit across Haiku/Sonnet/Opus calls. TITAN has no cross-agent spend tracking today. A lightweight PostToolUse hook logging (agent_id, session_id, model, input_tokens, cache_read_tokens, output_tokens, ET) to a local SQLite table would provide the per-agent spend dashboard currently missing, and enable circuit-breaker logic (block sessions exceeding 1M ET). Integration effort: low — one hook + one SQLite schema.
--permission-mode auto)Auto mode's two-stage server-side classifier handles safe/destructive action gating without human intervention. TITAN cron tasks run in bypassPermissions mode today (blanket trust). Switching low-risk tasks (email sends, newsletter fetches, asset index rebuilds) to --permission-mode auto tightens blast radius if a task goes rogue, while keeping bypassPermissions only for tasks that genuinely need it. Integration effort: low — settings.json + CLI flag changes per scheduled task.
The 2026 standard separates AGENTS.md from CLAUDE.md — AGENTS.md is scoped to multi-agent workflows, tool interoperability, and cross-agent contracts. TITAN packs everything into CLAUDE.md (approaching the 300-line performance threshold). Splitting into CLAUDE.md (universal rules) and AGENTS.md (SCOUT/VAULT/FORGE/GUIDE/ORACLE/DARWIN contracts, delegation rules, shared memory paths) reduces per-session token overhead and makes each agent's context more relevant. Integration effort: medium — restructure existing CLAUDE.md content; no new tooling.
/plugin MarketplaceClaude Code's official plugin marketplace (55+ curated plugins) is live via /plugin > Discover. TITAN hand-builds equivalent capabilities as custom skills. Auditing the marketplace for overlap with existing TITAN skills (SonarQube, Firecrawl, compounding-engineering-plugin) would identify where TITAN maintains code that's already supported and updated by the community. Integration effort: low — discovery cost only; installation is a single CLI command.
---
Real-time SAST, secret scanning (450+ patterns across 40+ languages), code quality gates, and dependency risk analysis directly in the Claude Code terminal. Reduces AI-introduced security regressions. Directly relevant for TITAN's Lambda/API gateway code and Silent Infinity app generation.
Install: claude plugin install sonarqube (or /plugin > Discover > SonarQube)
Why: TITAN has no automated security scan on generated code. SonarQube reports a 44% reduction in AI-related outages. Zero ongoing maintenance cost vs. building this capability as a custom skill.
Structured web scraping as a Claude Code skill — extracts clean markdown from any URL, handles JS-rendered pages, supports batch crawls. SCOUT currently falls back to WebFetch for raw HTML. Firecrawl returns structured data rather than raw HTML noise, saving context tokens on every research task.
Install: claude plugin install firecrawl (available via buildwithclaude hub and the official marketplace)
Why: SCOUT uses WebFetch as a fallback multiple times per deep-research session. Firecrawl cuts per-page token cost ~30% and handles dynamic pages that WebFetch cannot. Aligns with the Perplexity-first, WebFetch-as-last-resort token discipline already in TITAN's operating contract.
Official Anthropic plugin (2.3k GitHub stars) that enforces systematic learning loops: mistakes auto-update CLAUDE.md, confirmations auto-save as positive feedback, hooks trigger on correction patterns. TITAN manually delegates feedback saves to VAULT. This plugin automates the correction -> memory pipeline without requiring an explicit delegation step.
Install: claude plugin install compounding-engineering
Why: Directly replaces manual VAULT delegation for routine feedback saves. Every correction auto-propagates. Frees VAULT for higher-level memory consolidation and audit work.
---
Top agentic teams write a 5-15 line markdown PRD before invoking any skill, loaded as a resource rather than pasted into context. TITAN skills should accept a --prd flag that reads a lightweight intent file, rather than relying entirely on the skill's system prompt to infer scope. Prevents over-engineering when the task scope is narrow.
Professional workflows run long tasks in fresh subagent contexts with explicit handoff files (JSON summary of what was accomplished) rather than a single monolithic session. TITAN's symphony and newsletter skills are single-session today. Splitting into plan -> execute -> verify subagents with a shared state file reduces context drift and makes each phase independently restartable.
A PostToolUse hook (or subagent step) that runs a verification command and checks exit code before the agent marks a task complete. TITAN's operating contract says "verify, don't claim" but there is no enforcement mechanism. A hook that fires after every Write/Edit and runs a lightweight lint or test command catches regressions before the human sees them.
Rather than loading all domain rules at session start, expose TITAN's playbooks as MCP resources that Claude fetches only when a matching intent is detected. TITAN currently injects everything via CLAUDE.md. Moving app-specific rules, newsletter templates, and infra runbooks to MCP resources reduces base context size by an estimated 20-40%.
Route all Claude API calls through a lightweight local proxy that injects X-Agent-ID and X-Session-ID tags and logs ET values to SQLite. Adapted from the TrueFoundry pattern for local use with no cloud dependency. Enables per-agent spend visibility and the circuit-breaker pattern: block any session exceeding a configurable ET ceiling.
---
| Current TITAN Pattern | Replace With | Reason |
|---|---|---|
| Python subprocess in PostToolUse hooks | type: "mcp_tool" hook output (v2.1.118) | Native MCP dispatch, no process spawn overhead |
| Manual VAULT delegation for every feedback save | Compounding Engineering Plugin | Plugin automates correction -> CLAUDE.md loop |
| Polling loops in background cron tasks | Built-in Monitor tool | Monitor streams stdout; no sleep/recheck needed |
| Blanket bypassPermissions on all cron tasks | --permission-mode auto for low-risk tasks | Auto mode classifier provides safety gating at zero friction cost |
| All rules packed into CLAUDE.md | Split CLAUDE.md + AGENTS.md | 2026 standard; reduces per-session token overhead as TITAN scales |
| WebFetch for HTML extraction in SCOUT | Firecrawl plugin | Structured extraction, JS support, ~30% token savings per page |
---
Sources: Perplexity sonar-pro (2026-05-13), Perplexity sonar x4 (2026-05-13), daniellevantini.com, jls42.org, claudefa.st, firecrawl.dev, nimbalyst.com, truefoundry.com, github.blog, obviousworks.ch, usewire.io, checkmarx.com