Tokenomics of Agentic Security

Can the same AI agents you use to write code also do a good job of securing it? We ran a controlled benchmark to try to figure that out. 

The setup: the same 34 security prompts against 13 scopes with two agent configurations, where the only difference was the harness. One ran with Endor Labs' precomputed, deterministic evidence (reachability, call graphs, upgrade paths); the other worked from the raw repo and the open web.

The results:

• 91.7% fewer tokens (6.6M vs 79.5M)
• 2.8x faster completion
• 77.6% fewer tool calls (1,380 vs 6,165)
• Predictable cost (1.8x spread vs a 4x swing)

The unequipped agent used 12x the tokens, and almost all of the extra went to reconnaissance: grepping for version strings, walking manifests, re-reading the repo at every step. Its cost tracked the size of the codebase, not the prompt, swinging 4x between projects.

In this webinar we cover:

1. Why agents burn tokens: An agent re-sends its whole accumulated context at every reasoning step, so every repository read gets paid for again and again.
2. What the harness actually changes: Without evidence, the agent hand-rolls SCA from scratch, about 14 tool calls per prompt: 1,985 web searches, 849 greps, 783 manifest reads across the benchmark. Inside the harness it averages 3, because a lookup isn't a reasoning problem.
3. Confidently wrong, at 22x the cost: Asked for the highest-severity reachable finding, the agent without call-graph evidence spent 22x the tokens and returned a confident, well-cited, wrong answer, down to a fabricated version number.
4. Why the bill is hard to forecast: Without evidence, cost tracks your codebase, not your prompt. The same 34 prompts cost 2.7M tokens on a small repo and 10.5M on a large one. Inside the harness, the spread from cheapest to most expensive stayed inside a 1.8x band.
5. What it takes to run agents at scale: Deterministic evidence is the substrate agents reason from. It carries a cost of its own, though: on evidence-heavy tasks like upgrade impact analysis the win narrows to 72.5% of tokens and 27.8% of time, which is why token economics belong in your architecture review.

‍