The Agent Stack: Executive Summary

Scope: AgentVault, SmallChat, Stenographer, and Short-Hand — evaluated as a single ecosystem. Audience: stakeholders deciding whether/how to integrate these projects.

Sourcing note: This session's GitHub access is scoped to johnnyclem/agentvault only. AgentVault was evaluated directly from source. SmallChat, Stenographer, and Short-Hand were evaluated from their public GitHub READMEs and repository metadata (no source-level access). Treat claims about those three as directionally accurate, not verified against code.

What each project is, in one line

Project	One-line role	Language	License	Maturity
AgentVault	Deploys AI agents to Internet Computer canisters for persistent, 24/7, sovereign execution	TypeScript / Motoko	MIT	Active, 508 tests, v1.0 docs
SmallChat	Deterministic, in-process tool dispatch so agents pick the right tool without stuffing 50+ schemas into the prompt	TypeScript (+ Swift port)	MIT	Public, @smallchat/core on npm
Stenographer	A passive "court reporter" that tails agent conversation logs and builds a searchable GraphRAG index (entities, relations, decisions)	TypeScript	—	Alpha (0.1.0-alpha.2)
Short-Hand	Progressive, LSM-tree-style compaction of conversation history into a token-budgeted context frame	TypeScript	MIT	Public, npm package

The thesis

Long-running, autonomous agents need four things that a single chat session doesn't: a body that survives when the browser tab closes, reflexes that pick actions cheaply and reliably, a memory that records what happened, and a working-memory compressor that keeps that memory usable inside a fixed token budget. Each of these four repositories owns exactly one of those responsibilities:

 AgentVault        →  the body        (durable, on-chain execution + wallet + secrets)
 SmallChat         →  the reflexes    (deterministic tool selection, no schema bloat)
 Stenographer      →  the memory      (passive conversation observer + GraphRAG index)
 Short-Hand        →  working memory  (compacts raw history into an LLM-sized context frame)

None of them is useful alone for a fully autonomous agent — a body with no memory forgets everything on restart; a memory with no compaction blows the context window; reflexes with no body have nowhere durable to act. Together they describe a coherent, layered agent runtime, authored by the same person, with Short-Hand explicitly billed as "language middleware for Stenographer and SmallChat" — the seam between the memory layer and the reflex layer is intentional, not coincidental.

Key finding: the ecosystem is designed, but only partially wired

AgentVault today only integrates the SmallChat pattern — and even that is a from-scratch, in-repo reimplementation (src/orchestration/smallchat-*.ts), not a dependency on the published @smallchat/core package. Stenographer and Short-Hand are not referenced anywhere in AgentVault's source, tests, or docs. AgentVault already has two of its own, independently-built subsystems that overlap with what Stenographer and Short-Hand do:

• MemoryRepo (docs/memory-repo.md) — a git-style, on-chain versioned memory for structured state (commits, branches, merges), not conversational recall.
• Polytician enrichment (src/orchestration/polytician-enricher.ts) — pulls "concepts" from an external MCP server and stuffs them into the prompt with a hard character-count truncation, the exact naive-truncation problem Short-Hand's compaction engine was built to avoid.

So the four-project ecosystem is real as a design philosophy, but as of today it is one bridge built (AgentVault ↔ SmallChat, partially) and two bridges missing (AgentVault ↔ Stenographer, AgentVault ↔ Short-Hand).

Why this matters

• Cost & context efficiency: Polytician's truncate-at-N-characters enrichment is the weakest link in AgentVault's orchestration pipeline for any long-running session. Short-Hand exists specifically to solve this.
• Auditability: AgentVault already cares deeply about audit trails (VetKeys signing, on-chain commits, policy engine logging). Stenographer's decision-supersession/tombstone model is a natural fit for "why did the agent decide X" queries that MemoryRepo doesn't answer today.
• Duplication risk: Building deeper conversational memory into AgentVault from scratch would re-invent what Stenographer already does. The cheaper path is integration, not reimplementation (as was arguably already done once with SmallChat).
• Dependency risk: SmallChat, Stenographer, and Short-Hand are single-maintainer, low-star, alpha-to-early-stable projects. Depending on them directly (vs. vendoring, as AgentVault already does for SmallChat) trades duplication for supply-chain and API-stability risk.

Recommendation

1. Treat SmallChat's vendored pattern in AgentVault as the template: evaluate, don't blindly adopt — pull in Stenographer/Short-Hand behind the same kind of thin bridge/adapter AgentVault already builds for SmallChat and Polytician, so a breaking upstream change can't take down agent execution.
2. Pilot Short-Hand as a drop-in replacement for the truncation logic in polytician-enricher.ts — smallest surface area, clearest win.
3. Pilot Stenographer as a read-side companion to MemoryRepo for conversational Q&A ("what did we decide about X"), not a replacement for MemoryRepo's structured on-chain commits.
4. Defer adopting the published @smallchat/core package; the in-repo reimplementation is already tailored to Candid method dispatch and canister cycle constraints — re-platforming has a real cost and no clear benefit yet.

See docs/ecosystem/engineering-guide.md for the technical detail behind these recommendations.