Ch 4, The Runtime Loop Flashcards: The Claw AI Agent Ecosystem (Ava S. Barron)

Concept flashcards for Ch 4, The Runtime Loop, based on The Claw AI Agent Ecosystem by Ava S. Barron (ISBN 979-8-2526-0474-9).

Steps in one agent turn

Event received, context assembly, model call, response parsing, tool execution, follow-up model call, memory update, response delivery.

Response parsing branch

Text only: skip to memory update and delivery. Tool calls: execute, append results, call the model again, and repeat until no tools remain.

Session history default

The whole session history loads every turn; without pruning it grows unbounded and the user pays for all of it on each message.

Model-tool-memory sub-loop

Validate the call, execute the tool, append the result, call the model again; repeat until the response is text only.

Tool results as cost multiplier

Each appended result is billed again on every later model call within the same turn.

Tool call validation checks

Tool existence, schema conformance, permission tier, and rate limit. A failure returns to the model as a tool error result.

Limit of tool validation

It checks format and permission tier, not whether the outcome matches user intent; that is SOUL.md's and AGENTS.md's job.

Wrong tool selection

The most insidious tool-generation failure, since the call may succeed while doing the wrong thing. Prevent it with explicit do-not-use lines in SKILL.md.

End-of-session memory write

A model call that turns session history into a structured summary in memory/YYYY-MM-DD.md; it can be routed to a cheaper model.

What not to log

Verbatim transcripts, facts already in USER.md or MEMORY.md, credentials, full private documents, and unconfirmed speculation.

Memory security boundary

Memory files are plaintext readable by anyone with filesystem access. Log that a communication happened and its action items, not its content.

Why promote to MEMORY.md

Only today's and yesterday's logs load by default, so durable facts must move to MEMORY.md to stay in context.

Session

Stored state with history, start time, owner (user plus channel), and memory log path. Kept in SQLite by default; survives restarts.

Session isolation key

By default sessions are separate per channel and per peer, which prevents personal and work context from bleeding together.

Summarize vs truncate

Summarize the oldest turns at the soft threshold to keep continuity; truncate at the hard threshold as a backstop that drops them.

Heartbeat benefits

Proactive task triggering, system health checks, advancing background tasks, and memory maintenance.

Heartbeat model choice

Almost always a local or cheap model; typical heartbeat work fits a 1B-3B parameter local model.

Cache-friendly context ordering

Load stable content first, keep timestamps out of the system context, and keep tool schema order fixed.

cacheControlTtl

How long cached blocks persist; default 300 seconds. Raise it when interactions are hourly or less frequent.

Silent memory write failure

For example, disk quota exceeded: the reply looks normal but nothing persists. Monitor workspace disk and alert at 80%.

Context accumulation problem

Long multi-tool tasks pile results into one turn's context. Mitigate by compressing results, limiting verbosity, or splitting into separate turns.

Task checkpointing

For tasks over five tool calls, track the plan and progress in tasks_in_progress.md so a restart can resume instead of losing work.

Graceful shutdown

Enable maintenance mode, wait for in-progress turns to drain, then stop the container.

maxTurnDurationSeconds

Per-turn timeout so a hanging tool call can't block the queue forever. Set it near the 99th percentile of legitimate turn length.

Identity file reload

Identity files load at session start, so edits to SOUL.md or AGENTS.md during a session need a session restart.