Agent Dispatcher — Automatically routes tasks to the right role, skills, tools, and context

So most people know now, If you properly configure your AGENTS and SKILLS files, you can get massive gains in how an AI agent performs. But the opposite is also true. More agents, skills, and instructions don’t automatically mean better results. They can introduce context bloat, overlapping instructions, bad routing, unnecessary tool access, and conflicting guidance. Updates to the underlying agent/model can also change how well an existing setup behaves. That’s why I built Agent Dispatcher . Instead of loading everything into every task, it tries to assemble the smallest trustworthy capability stack for the requested outcome . Initially, it includes: 27 roles · 79 local skills · 31 external skills · 8 recipes · 19 MCP servers · 50 detection signals Everything is extensible, so you can add your own roles, skills, MCPs, recipes, and routing logic. How it works Your request → specialist role → context plan → execution → verification → evidence/result The context plan determines the relevant: skills project files tools/MCPs guidance verification requirements A role owns the outcome. A skill supplies a reusable method. An MCP/tool provides a capability. A recipe suggests a workflow across roles. The context plan assembles what the specialist actually needs before planning the work. One important design rule: skills should generally be narrower than agents. Good skills: systematic-debugging browser-verification database-migrations accessibility api-integration Avoid skills that basically recreate an entire engineer/agent role. The goal isn’t to build the largest possible agent or skill library. It’s to load less, but load the right things . The dispatcher also defines what evidence should count as “done.” Routing to the right specialist isn’t very useful if there’s no way to verify the result. These are instructions and validation rules, not a sandbox. Authorization still comes from the user and the host’s permission controls. Detecting a stack, selecting a tool, or switching roles does not grant permission to use it. Routing evals I also added optional Jev integration and tested it against the built-in routing approaches. On this particular 162-case routing benchmark, Claude/Codex-style routing performed substantially better, so Jev remains optional rather than being the default. Metric Keyword Jev Claude Agent top-1 23/162 143/162 158/162 Acceptable route 29/162 154/162 162/162 Obvious top-1 9/54 52/54 54/54 Near-neighbour top-1 6/54 49/54 53/54 Ambiguous top-1 3/27 17/27 25/27 Negative top-1 5/27 25/27 26/27 Skill precision / recall .31 / .56 .68 / .71 .74 / .90 Tool precision / recall .15 / .23 .65 / .97 .69 / .95 The keyword baseline is useful here because it shows why simple keyword → agent routing starts breaking down once the capability library gets larger. Repo: github.com/nahid-sparktales/agent-dispatcher I have more updates planned, but I’d especially like feedback from people running larger agent/skill setups. If you’ve run into routing failures, context bloat, conflicting skills, or have ideas for additional roles/skills/verification strategies, LMK.