TokenEconomizer cuts coding agent costs by 18-31%

Microsoft Research and UC Berkeley released TokenEconomizer, an open-source plugin that learns from earlier task runs in a repository to reduce token waste in later ones. The tool works as an external controller that does not modify agent code or model weights. It supports Claude Code, Codex, GitHub Copilot CLI, and Pi.

What changed

TokenEconomizer addresses a specific cost problem in coding agents. When researchers profiled 60 SWE-Bench tasks, they identified three recurring sources of waste. First, agents repeatedly call the same tools across sessions: once five tasks precede a new one, roughly 30% of tool calls repeat an earlier action. Second, agents load unused schemas on every turn: of 23 tool schemas that load automatically, 16 are never called but 3 cover 98% of actual calls. Third, failed runs consume disproportionate tokens: while failed tasks represented 17% of the total, they accounted for 32% of overall token spend.

TokenEconomizer intercepts agent decisions at three points. The micro-skills mechanism captures one-line facts like a working test command from completed traces, then inserts them into future prompts after multiple confirmations. Tool pruning disables schemas that go unused for three consecutive tasks, cutting prompt size significantly. On Django with Claude Code, disabling unused tools reduced the first task request from 43,351 to 16,295 tokens. Failure-loop intervention sets a turn budget based on successful run patterns and halts agents when they repeat errors.

On 277 SWE-Bench tasks across five agent-model pairs, TokenEconomizer lowered mean token cost by 18.4% to 30.9%. The strongest result came from Claude Code with Opus 4.8, achieving 30.9% cost reduction, which rose to 33.2% when combined with another cost-control method called RTK. The plugin showed cost savings across all 30 repository-setting combinations tested. In comparison, fixed alternatives like RTK and RepoMap individually raised costs in 12 and 15 cells respectively.

The package installs as a Python module with agent-specific plugins and requires Python 3.10 or newer. No GPU or dedicated API key is needed. Learned data stays local on the user’s machine. Tool pruning and failure-loop intervention are enabled by default. Micro-skills are disabled by default because the mechanism that generates facts adds additional model calls.

Why it matters

Coding agent deployment has become expensive for organizations. Tool pruning alone saves 13.2% to 17.7% of Claude Code token costs by disabling unused schemas, requiring no additional model calls and running by default. Micro-skills save 11.8% to 14.3%, and the failure-loop guard contributes 5.4% to 7.1%.

The learning-based approach matters because it adapts to each repository rather than applying fixed strategies everywhere. Agents waste tokens by reprocessing the same information across sessions: the system prompt, full history, and all loaded tool schemas get billed on every turn. The three mechanisms target these specific inefficiencies. By capturing what worked before, disabling tools that never get used, and preventing agents from repeating error patterns, the system learns local patterns that fixed optimization strategies cannot capture.

Results also showed that savings accumulate over time. Most repositories cost more on early tasks but all were cheaper by the sixth task, suggesting organizations should measure cost trends over time rather than expect immediate returns. Outside SWE-Bench, costs fell 16.0% to 21.2% on Web-Bench and 16.5% to 19.6% on CredData. ReposVul security audit tasks showed only 6.4% to 9.3% savings because shorter runs leave fewer optimization opportunities.

Pass@1 (the rate at which agents solve tasks on the first attempt) improved in four of five settings. Only Claude Code with Opus 4.8 saw a 1.0-point drop, meaning the tool rarely worsens task completion rates while cutting costs. This stability is important because cost savings that come at the expense of fewer successful runs create a trade-off that may not justify adoption.

What to test

Before deploying TokenEconomizer across repositories, teams should verify several variables in their own environments:

• Warmup period: Most repositories cost more on early tasks but all showed cost reductions by the sixth task. Early deployments may not show immediate savings. Measure cost trends over time to see the full benefit rather than judging on day one.

• Repository variability: Results vary significantly by codebase type. Short tasks like security audits saw only 6.4% to 9.3% savings because less code means fewer opportunities for optimization. Test TokenEconomizer on actual repositories before committing resources.

• Tool pruning impact: Tool pruning delivers the largest single benefit and runs by default. It requires no additional model calls, making it low-risk to test first. Verify the token reduction on your repositories aligns with the 13.2% to 17.7% range on Claude Code.

• Micro-skills trade-off: Micro-skills are disabled by default because the LLM proposer that generates facts adds model calls. Enable only if the added calls do not exceed expected token savings. Test whether facts like working test commands actually help your agent’s task success.

• Failure-loop intervention: This mechanism halts agents after one failed redirect if the loop recurs. Confirm that this approach aligns with your team’s desired behavior for stuck agents rather than other retry strategies.

• Agent support: TokenEconomizer works only with Claude Code, Codex, Copilot CLI, or Pi. Verify your team uses one of these before adopting.

• Pass@1 stability: While task success rates held steady or improved on most models, Claude Code with Opus 4.8 saw a 1.0-point drop. Confirm that any small loss in completion rate is acceptable for your use case.

The conclusion

TokenEconomizer addresses a concrete problem: coding agents generate recurring waste that fixed optimization strategies cannot eliminate. The 18.4% to 30.9% cost reduction on SWE-Bench is substantial, especially given that pass rates held steady or improved on most configurations.

The plugin is free, Apache 2.0 licensed, and requires no changes to existing agent architecture. Savings build over time as the system learns from completed tasks, so early adoption will not yield immediate returns. Results vary by repository and task type, meaning teams should test on their own codebases rather than assume benchmark gains will transfer directly.

For organizations deploying agents frequently across multiple repositories, the learning-based approach makes sense because it adapts to local code patterns. Watch whether savings sustain and scale as more tasks run, and whether results hold on codebases outside public benchmarks. If your team already tracks agent token spend as a cost center, TokenEconomizer deserves a trial on high-volume repositories first.

Related reading

• AKV cache engineering cuts agent costs 27.7%
• agent-desktop: Open-source agent computer use
• Agensh: Decentralized multi-agent harness for 1,024 agents