Budget & identity
Two Talon agent identities use separate spending limits. Admission checks estimates before provider dispatch; they do not guarantee a final invoice ceiling.
Use case · Engineering and AI platform
Govern coding agents without pretending to optimize their code.
Coding agents can consume model budgets, call company tools and run third-party optimization plugins. Talon governs the model/provider and MCP paths routed through it, attributes use-case cost and records verifiable decisions. The coding agent, its plugin and the local runtime remain separate systems.
EXPERIMENT: coding task stream
baseline → Talon agent A
optimizer → Talon agent B
policy separate $5 estimate budgets
proof signed model-call records
quality independently graded tasks
conclusion optimizer savings NOT provenCoding tools retainedSeparate agent identitiesCost-admission boundariesMCP permissionsSigned evidence
Proof 1 · TokenEconomizer × Talon
Can we independently assess a coding-agent optimization claim?
Microsoft's TokenEconomizer is an external research/plugin system, not a Talon feature. The merged governed TokenEconomizer experiment compares independently authenticated baseline and optimizer arms, each routed through released Talon v1.15.3 with estimate-based budgets, model-call attribution and signed receipts. Correctness grading belongs to the experiment; Talon contributes control and evidence.
Quality & cost
The study accounts for attempted tasks, independent grading, provider calls, cache-aware costs, receipt integrity and truncated runs without quietly counting unreached tasks as failures.
Third-party plugins stay independent
The TokenEconomizer code, pruning logic and any memory/learning remain external. Talon does not automatically optimize, prune tools or learn micro-skills.
Actual result, not a product claim
Both arms completed 24/24 tickets. No optimization gain was demonstrated.
In the October 23 internally authored 24-ticket study, both arms passed all 24 tickets and Talon-recorded estimated spend was $3.2904 baseline vs $3.3560 with the plugin—about 1.99% more in the plugin arm. The plugin logged 11 pruning proposals but zero applied pruning, injected facts or failure-loop nudges. One seed of a synthetic workload cannot establish plugin efficacy.
A separate official 20-task Web-Bench pilot stopped early at unequal task counts: 15 baseline tasks and 10 experimental tasks were reached. The nine commonly completed tasks with complete usage showed fewer effective tokens on the experimental arm, but this selected subset must not be presented as full-benchmark savings or reproduced research results. Several failures' root causes remained unestablished.
What Talon did prove: separately governed provider access and budget-admission decisions with verified signed evidence. The study did not prove token reduction, automated skill learning or broad coding-quality conclusions.
Proof 2 · Copilot release tools
Can a coding assistant inspect a release without publishing it?
A separate demo keeps the real GitHub Copilot CLI. Supported model traffic and MCP tools cross Talon under one authenticated coding use case. A synthetic release-status and preparation workflow is allowed, while the publish operation does not reach the upstream release server; a separately controlled negative test exercises pre-dispatch denial.
Local shell commands, file editing and direct Git pushes that bypass Talon are out of scope. The client's own local permissions are a separate boundary, not a Talon security guarantee.
coding assistant
→ Talon model gateway
→ Talon MCP proxy
release_status allowed
release_prepare allowed
release_publish NOT upstream
local shell / filesystem
→ outside this Talon proofBuyer evaluation
What a platform engineer should verify
Attribution
Do model calls belong to the authenticated coding use case? Do asserted sessions and subagent labels remain distinct from credential identity?
Prevention
Can a disallowed provider or intercepted MCP operation be stopped before dispatch? Is the upstream call count truly zero on denial?
Evidence
Can a reviewer check the reason, estimated budget, completed/unknown outcome and signed record? Are plugin costs and quality comparisons reported without selection bias?
No universal local-agent monitoring, automated loop prevention, or exactly-once external effects are claimed. All numbers above describe bounded demo studies, not customer production workloads.
Reproduce before committing
One coding agent, one action boundary, one verifiable receipt.
Start with the completed evidence-driven study or the Copilot release-control demonstration. Verify your own supported interception path before treating it as coverage for all developer tools.