No public counterpart in snapshot
Entry: Why the Entry Point Became Simpler
No Direct/Research decision, authorization catalog presentation or ordinary follow-up handling.
Skip to contentETF AGENT / NOTES Build small.
Source lineage ↗ Full catalog →
THE OPEN ETF AGENT FIELD GUIDE
Build small.
Keep evidence.
From engineering essays to a research Agent you can run. One module, one inspectable increment at a time.
33module essays in the reading map
31rows in the coverage snapshot
15bounded execution targets
0.11work in progress · not stable
01 / THE BUILD LOG
One increment. A clearer system.
Fixed development snapshots, not a release registry. Open any stage to inspect its original code; each new stage has one distinct delivery focus.
- 0.1First readable reportRuntime · Gateway · Tool Host↗
- 0.2Methods meet toolsSkill · MCP · SSE↗
- 0.3Cooperate within boundsChild · Hook · shared budget↗
- 0.4Keep the originalResult · Context · Retrieval↗
- 0.5Notes are not factsCompaction · explicit Memory↗
- 0.6Deliver by identityArtifact · Revision · Reader↗
- 0.7Contribute and recoverPlugin · SQLite · lease / fence↗
- 0.8Commit the answer bodyPublication · revision CAS↗
- 0.9Continue the same RunSubmit · Worker · readback↗
- 0.10One result, actual consumersResult / Reference · report · CSV↗
02 / THE READING MAP
Read the idea. Inspect the boundary.
Start here: the first report you can read back →
Curated through the existing module catalog, not a second article index. A receipt proves only the named slice. Broader essays and later additions remain distinct from implemented coverage.
01Research, reading & computation
No public counterpart in snapshot
Web Research: From Search to Deep Reading ↗
No HTTP acquisition, continuous reading, page state or source revocation.
No public counterpart in snapshot
When an Agent Says Today: User Time and Market Sessions ↗
No trusted timezone, calendar coverage or market-data dataAsOf.
No public counterpart in snapshot
PDF: From Text to Evidence ↗
No PDF text, page-image, layout-location or multimodal reading.
Limited target · receipt available
Subagents: Isolation and Specialized Collaboration ↗
One-level scripted Child shares budget and returns original results; not live-model experts.
No public counterpart in snapshot
A Custom Expert Should Be More Than a Prompt Field ↗
No expert CRUD, frozen versions, catalog permissions, historical identity or recovery.
No public counterpart in snapshot
Adaptive Research and Python ↗
No model-driven research-question revision, arbitrary authorized Python or data-artifact handoff.
No public counterpart in snapshot
Python: From Tool to Computational Workbench ↗
No isolated execution, code/input identity or artifact retention after failure.
No public counterpart in snapshot
Connecting Portfolio Data Through a Read-Only MCP ↗
No portfolio-report, share/cash/snapshot semantics or proposal calculation.
No public counterpart in snapshot
Controlled Refresh for Stale Data ↗
No refresh-task identity, scope/fingerprint deduplication, source refusal or atomic publication of new evidence.
02Context, memory & knowledge
Limited target · receipt available
Externalized Results and Input Completeness ↗
Lexical candidates → actual page reads → exact original rereading; not a complete original-source coverage platform.
Limited target · receipt available
Main Context Compaction Without Rewriting Facts ↗
Derived notes remain separate from originals and support rereading after adoption; summary quality unreviewed.
Not yet assessed
Cross-Run Compaction: Why an Agent Cannot Simply Summarize History ↗
Not included in this coverage snapshot. No execution claim.
Limited target · receipt available
Memory: From Saving and Recall to Safe Forgetting ↗
Save/revise/forget using explicit synthetic intent; not an authenticated user Memory service.
No public counterpart in snapshot
RAG: Revocation and Historical Replay ↗
No dynamic ACL/revocation propagation, publication eligibility or replay reauthorization.
03Runtime, extensions & observation
Limited target · receipt available
Provider: Why a Unified API Still Needs Shared Admission ↗
Gateway encoding and matching via offline transport; no shared Provider admission.
No public counterpart in snapshot
Why the Next Model Request Still Queues After HTTP 200 ↗
No cross-process queue, lease/isolation, remote unknown/release or PostgreSQL control database.
Limited target · receipt available
Extension Host: From Feature Phone to Smartphone ↗
Trusted packages use the same Host; not an isolation sandbox or complete multi-ecosystem Host.
Limited target · receipt available
MCP Host: From Discovery to Usable Tools ↗
Real stdio synthetic MCP discovery, admission and acquisition; not a financial-data service.
Limited target · receipt available
Tool Surface: Why the Model Cannot See an Available Tool ↗
Child advertises only authorized read tools; not complete dynamic capability orchestration.
Limited target · receipt available
Prompt Injection Is More Than Malicious Text: An Agent Content Trust Boundary ↗
Context rejects fabricated original-source identity; not the complete A35 security module.
Limited target · receipt available
Multi-Ecosystem Hooks: Claude, Pi and Codex ↗
The same Host revalidates Hook argument changes and preserves the original call; no multi-ecosystem translation layer.
Limited target · receipt available
Observing the Tool-to-Model Input Lifecycle ↗
Synthetic workers acquire sources through existing Runtime/Host; not a complete observability platform.
Limited target · receipt available
Run Budget and Constraint Governance ↗
Root-budget refusal prevents the next model request; not every budget policy.
No public counterpart in snapshot
Agent Usage: Requests, Tokens and Shadow Credits ↗
No account/physical-attempt identity, attribution, price catalog or shadow cost accounting.
Limited target · receipt available
Why an Agent Cannot Execute a Stop Command Immediately ↗
Durable cancellation prevents late-proposal dispatch across processes; not a complete control-event system.
Limited target · receipt available
Scheduled Research Beyond a Conversation ↗
Explicit submit and separate work process read back the original Run; no schedule trigger, Task orchestration, notification or foreground concurrency.
No public counterpart in snapshot
Langfuse: From Traces to Trustworthy Observability ↗
No Langfuse exporter/span/session ownership or remote-failure isolation.
04Control, publication & delivery
Limited target · receipt available
Publication: Why Answers Need Their Own Delivery System ↗
The same Runtime commits body text for independent CLI readback, separate from report Artifact; no Web/SSE/Worker or complete publishing system.
No public counterpart in snapshot
Keeping Progressive Answers Coherent ↗
No committed-block/revision system, progressive frontend display or history consistency.
Not yet assessed
When a User Dislikes an Answer: Feedback Is Not Automatic Learning ↗
Not included in this coverage snapshot. No execution claim.
No public counterpart in snapshot
Async Research: Delivering on Time After Users Leave ↗
No continuing task after leaving, by/at delivery agreement, notification or same-Run recovery.
No public counterpart in snapshot
After an Agent Asks: Durable Human Response Adoption ↗
No request/response/adoption/effect identity, Human Park/Continue or authenticated user reply.