Pimsy LLM

Multi-Agent Orchestration

Preview

Pimsy delegates to specialized workers while remaining the single point of contact.

Hard objectives are decomposed across specialized sub-agents. The user does not manage a fleet — they talk to Pimsy, which orchestrates, synthesizes and remains accountable for the result.

Delegation topology
                    user objective
                          │
                          ▼
                      ┌────────┐
                      │ PIMSY  │  orchestrator
                      └───┬────┘
                          │  task decomposition
      ┌──────────┬────────┼────────┬──────────┐
      ▼          ▼        ▼        ▼          ▼
  research   engineer  analyst  verifier  security
   agent      agent     agent    agent     review
      │          │        │        │          │
      └──────────┴────────┼────────┴──────────┘
                          ▼
                   Pimsy synthesis
                          ▼
                     verification
                          ▼
                    final outcome

Sub-agent contract#

  • A sub-agent receives a narrowed objective, a subset of the parent capability set and its own budget.
  • It cannot escalate privileges, spawn peers outside its allowance, or write to long-term memory directly.
  • Its output returns to the orchestrator as an observation with provenance, not as a final answer to the user.
  • Conflicting sub-agent conclusions are surfaced to the orchestrator for explicit resolution.

Specializations#

AgentTypical capabilitiesProduces
Researchresearch.webEvidence sets with citations
Engineercode.execute, files.writePatches, tests, running services
Analystcode.execute, dataset accessModels, tables, quantified findings
Verifierread-onlyPass/fail findings and counter-arguments
Security reviewerread-onlyRisk findings against the produced change
Operatorcomputer.browserCompleted GUI workflows and evidence screenshots
ts
await pimsy.tasks.create({
  objective: class="tok-str">"Evaluate three deployment strategies for our inference fleet and recommend one.",
  orchestration: {
    maxConcurrentAgents: class="tok-num">4,
    roles: [class="tok-str">"research", class="tok-str">"analyst", class="tok-str">"verifier"],
    synthesis: class="tok-str">"orchestrator"
  },
  budget: { steps: class="tok-num">120, wallClockSeconds: class="tok-num">3600 }
});

Last updated 2026-09-11