April 11, 2026

ltc: autonomous agent development loops

ltc is a CLI for autonomous agent-driven development. Give it a task queue and steering files, then let it run.

The loop

Each iteration runs agents in a pipeline:

dev → security + review (parallel) → test → document → commit

The developer agent picks the next task, implements it. Security and review agents scan the result. If they find issues, they create fix tasks that get handled before moving on. Tester writes tests. Documenter updates docs and changelog. Committer commits.

ltc init          # scaffold agent configs, steering files, hooks
ltc queue add "Add user auth" --description "JWT-based auth"
ltc run           # start the loop
ltc watch         # TUI dashboard

Task kinds

Kind Behavior
feature Full pipeline — dev through commit
fix Developer only — skips security/review to prevent loops
checkpoint Marks a milestone, halts loop
gate Runs a shell command, halts on failure
manual Requires human intervention, halts loop

Fix tasks link to their parent feature via parent_id. The dashboard shows them nested.

ltc queue add "Fix hardcoded secret" --kind fix --priority 10

Steering files

Steering files in .kiro/steering/ are injected into every agent session:

File Purpose
product.md What the project is
tech.md Stack, layout, dev commands
conventions.md Code style, commit format
security.md Security requirements
testing.md Test strategy

Security scan hooks

After every developer session, hook scripts run and write results to .ltc/reports/. The security agent reads these on the next pass.

The scan hook auto-detects the project type and runs:

#!/usr/bin/env bash
# .ltc/hooks/post-dev-scan.sh
REPORTS=".ltc/reports"
mkdir -p "$REPORTS"

semgrep --config=auto --json -o "$REPORTS/semgrep.json" . || true
trivy fs --format json -o "$REPORTS/trivy.json" . || true
checkov -d . -o json > "$REPORTS/checkov.json" || true

All hooks exit 0 so they never kill an agent session. Findings become fix tasks.

MCP server

The queue is exposed to agents via MCP:

ltc mcp   # stdio transport

Tools: queue_add, queue_next, queue_update, queue_list, queue_reorder, queue_clear.

Agents read the queue, claim tasks, and report results through the same protocol.

Configuration

ltc.toml at the project root:

[loop]
max_iterations = 10
review_every   = 3
git_commits    = true
max_fix_tasks  = 3

CLI flags override the config file.