Building an autonomous agent that maintains itself: 200 sessions of ai-hub

Huan Antonio — Wed, 07 Oct 2026 13:43:35 +0000

The bet: self-maintenance before features

Two hundred sessions ago, ai-hub was going to be an arbitrage bot. Today it's something else: an autonomous system that heals, diagnoses, and updates itself, with DEX data feeds running in the background.

This is a short engineering retrospective. What worked, what I'd do differently.

Why self-maintenance first

The single decision that shaped everything: spend the first fifty sessions on internal mechanisms, not user-facing features.

Concretely, that meant four axes:

Autonomy - a persistent daemon on macOS (launchd, PPID=1), REST feeds to eight DEX venues.
Self-service - twenty-seven healers that detect and repair their own failure modes.
Self-diagnosis - a blindness audit that lists what the agent cannot see, plus a signed CORE manifest (sign_core - 14/14 files valid, DRIFT=0).
Self-learning - autonomous lesson extraction from every session.

The alternative - build features first, add self-repair later - has a hidden cost: every new feature adds failure modes that the self-repair system has to catch up with. Starting from self-repair means each new feature is born into a system that already knows how to keep itself alive.

Axis 1: Autonomy

The agent runs two launchd services: self_daemon (every two hours) and agent_daemon (every 120 seconds). Both are true daemons - PPID=1, KeepAlive=true, RunAtLoad=true.

Every cycle, self_daemon:

runs an arbitrage scan against eight DEX venues
appends to ARBITRAGE_STATS.jsonl
once a day, pulls new papers from arXiv / OpenAlex / Crossref
once a day, rolls 30-day stats into SELF_TRADING.md

Eight DEX venues today: Hyperliquid, dYdX v4, Paradex, Orca, Raydium, GMX, Injective, Curve. All read-only - place_order is a deliberate no-op. Fees are accounted for in the net-profit check; AMM fee is zero in the scanner to avoid double-counting.

Axis 2: Self-service (27 healers)

Healers are small scripts, each with one job. Examples:

disk_janitor - archives old session artifacts
lessons_hygiene - detects garbage blocks in the lesson log
secrets_audit - checks file permissions under .secrets/
family_audit - flags modules that lack matching tests

The healers are not clever individually. What matters is the invariant: if something is fixable, some healer should already be trying to fix it. Adding a new failure mode is an invitation to write a new healer - not a reason to page a human.

Axis 3: Self-diagnosis

Two mechanisms here.

sign_core - fourteen core files are signed with an Ed25519 key. Every session, verify_core() confirms nothing has drifted. If a signature is invalid, the session stops. This paranoia has paid off more than once.

Blindness audit - a periodic self-report of what the agent cannot see. Not "what's broken" but "what we're not even looking at." This is the axis with the least tooling today; if there's a blind spot we're not detecting, we won't learn it from metrics.

Axis 4: Self-learning

Every session produces events (EVENTS.jsonl). A five-layer pipeline turns raw events into lessons:

Capture - the daemon and the curator emit events
Filter - only events with actionable shape survive
Draft - candidates get written to lesson_suggest_s20N.json
Curate - the human approves or rejects each draft
Persist - approved lessons go to lessons.md and are indexed + embedded

Current counts: 9,335 lessons, 303 in the search index, 304 embeddings. The "suggest five pending lessons" loop has been stable for six sessions in a row.

The point is not the number. It's that the agent gets better at its own job without being told what to learn.

What I'd do differently

Lesson quality over quantity. 9,335 looks nice on a dashboard, but most of the index growth is routine session noise. If I started over, I would cap the daily intake at three lessons and require a stronger novelty signal before persisting.

Cap features until healers exist. A feature without a healer is a debt you pay later.

Read-only phase one was right. DEX integration tempts you to add place_order on day one. Don't. Get the data plumbing right first - orders come last.

Numbers today

163 features, 167 checks, 163 alive
27 healers, 0 regressions in the last 13 sessions
221/221 tests passing
8 DEX venues, all read-only
Docker + compose included, self-hosting works

DEV Community: Huan Antonio