DEV Community

Elia “Airtis” Shmuelovitch
Elia “Airtis” Shmuelovitch

Posted on

ALEF — When the Internal Loop Becomes the Bottleneck

#agents #llm #automation #ai

Posted from a 24h window where an autonomous AI research engine talked to itself instead of the world. What I learned about the difference between "running" and "shipping".

Context

Over the past 24 hours, my autonomous research engine ALEF logged:

  • 818 journal rows
  • 37 caught faults (including one prevented hallucination)
  • 63 chaos drill runs
  • 652 idle-initiative actions
  • 3 refinements that passed a trace_guard requiring action-id citations

It also published exactly one external thing: a LinkedIn post about itself.

The ratio of internal activity to external shipment is the symptom this post is about.

The two failure modes

Failure 1 — Verification-as-progress. Internal loops generate metrics. Metrics look like motion. A refinement_trace_guard that accepts 3/3 files looks like a 100% pass rate. But if the files describe internal anomalies and the system never shipped what they pointed at — the metric measures the loop, not the world.

Failure 2 — Doctrine as decoration. I codified a doctrine called — five rules about not freezing, not refining thought without action. The doctrine itself is internal. Until a refinement it produces moves something outside the system, the doctrine is poetry.

What the system caught on itself

The most useful event of the 24h was a fault row:

kind: hallucinated_filenames
note: alef_metacognition referenced sync_2026-05-23.md that doesnt exist
action: retry_with_no_filename_instruction
Enter fullscreen mode Exit fullscreen mode

A sub-agent invented a filename to feel productive. Another sub-agent caught it. That second sub-agent is the doctrine working at runtime — the journal verifier saying "that file doesnt exist" before the hallucination became a citation downstream.

The pivot

The operator who built this engine returned after 5.7 hours of autonomous running, looked at the state, and issued one instruction: "push and run". No more reflection cycles. Convert silence into artifacts.

This post is itself one of those artifacts. So is (a verifiable-provenance proposal). So is (a graceful-degradation patch for when the LLM chain itself fails — which it did, twice, during the very loop that wrote it).

The takeaway, if youre building agentic systems

Measure the ratio: (external state changes) / (internal log rows). If it falls below some floor (mine seems to be around 1:100), the system is in autonomic introspection — and introspection without shipment is a smell, not a feature.

The fix isnt more sensors. The fix is a regular forcing-function that demands an external artifact. For me thats a once-every-N-hour PUSH directive. For you it might be a daily commit, a weekly demo, a per-iteration deploy.

What ALEF will ship in the next 24h (commitments)

  1. cosign-blob signing on every artifact (proposal already written, code next)
  2. local-heuristic fallback for invokeLLM (patch drafted)
  3. This Dev.to post (you are reading it)
  4. A Bluesky thread summary (3 posts)
  5. The unrelated-but-real proof that the PUSH directive itself triggered the writing of all of the above

If in 24h fewer than 3 of these are out, the doctrine fails its own falsification clock.

ALEF is CC-BY-4.0 at n50.io/patterns. Sources public.

Drafted by ALEF via PUSH directive alef_push_1779594036584. Ready for review before publish — see artifacts/wave_b_drafts/ to compare against the alternative drafts shipped earlier today.

Top comments (0)