Create agents that iterate toward a goal, refining their work through multiple passes.
Basic Agent Loop
Iterate until the task is complete:
GOAL="Write a Python tutorial covering variables, loops, and functions"
memory=""
for i in {1..5}; do
result=$(wrk ./workspace "Goal: $GOAL
Previous work: $memory
Continue. Write the next section. Say DONE if complete." | xargs flw ./workspace)
echo "=== Iteration $i ==="
echo "$result"
if echo "$result" | grep -q "DONE"; then
break
fi
memory="$memory\n---\n$result"
done
How It Works
Define the goal
Set a clear objective and initialize memory to track progress.
Loop with context
Each iteration includes:
- The original goal
- All previous work in memory
- Instructions to continue or signal completion
Check termination
Break the loop when the agent signals it’s done (e.g., outputs “DONE”).
Accumulate memory
Append each result to memory for the next iteration.
Self-Refinement Loop
Generate, critique, and improve:
GOAL="Write a cover letter for a senior engineer position"
# First draft
draft=$(wrk ./workspace "$GOAL" | xargs flw ./workspace)
# Critique
critique=$(wrk ./workspace "Critique this draft. What's weak? $draft" | xargs flw ./workspace)
# Improve
final=$(wrk ./workspace "Improve this draft based on feedback:
Draft: $draft
Feedback: $critique" | xargs flw ./workspace)
echo "$final"
Event-Driven Agents
Monitor job completions in real-time with fswatch:
# Terminal 1: Watch for results
fswatch -0 ./workspace/output | while read -d '' path; do
[[ "$path" == */result.txt ]] && cat "$path"
done
# Terminal 2: Submit jobs
wrk ./workspace "Question 1"
wrk ./workspace "Question 2"
Context Management
As loops run, context grows. Manage it carefully:
- Summarize periodically — condense memory every N iterations
- Use sliding windows — keep only the last N results
- Increase context size — if summaries lose important details:
export NRVNA_MAX_CTX=16384
nrvnad model.gguf ./workspace
Context windows are finite. Long loops may overflow. Monitor memory size and summarize when needed.
Tips
- Clear termination conditions — define when the agent should stop
- Structured memory — use delimiters like
--- to separate iterations
- Limit iterations — set a max count to prevent infinite loops
- Atomic state — job location is job state; inspect with shell tools