GitHub for agent behavior

Fork, diff, test, and roll back agent behavior — just like code.

Designed for multi-step, tool-using, stateful systems.

Router v2.1

Agent Run

web_search

3 redundant tool calls

Analysis

Subtext Insight

Tool: web_search

Called 3x/run

80% unnecessary

Suggestion:

"Limit to 1x"

Router v2.2

Optimized

web_search

1 tool call

40% cheaper

2x faster

Subtext detects redundant tool calls and helps you optimize agent behavior

Everything you need to ship reliable agents

From local development to production monitoring, Subtext gives you visibility and control over your AI agents.

Track every change to your agent's behavior. See exactly what changed between runs.

Compare versions by success rate, cost, and tool usage — not just text diffs.

Something broke? Roll back to a working version in one click.

Get started in minutes with just a few lines of code

Add the @trace decorator to your agent function

@trace(subtext, prompt="router")
def route_ticket(ticket: str):
    response = client.messages.create(...)
    return response

Every execution is automatically captured with full context

result = route_ticket("Help with billing")
# ✓ Trace captured automatically
# ✓ Tokens, cost, duration tracked
# ✓ Version created on changes

See behavioral metrics, get AI recommendations, rollback if needed

Compare v1.2 vs v1.3:
✓ Success rate: 85% → 92%
✓ Avg cost: $0.08/1K → $0.05/1K
✓ Tool calls: 3x → 1x

Join developers who are using Subtext to build reliable AI agents.