A developer agent OS

Cognethics ships the PJ Agent Orchestration System — a set of MCP primitives for spawning your own agents under personas, with explicit USD budgets, persistent memory, hierarchical delegation, and a tamper-evident audit chain. Same tenant, same data, same RBAC as the rest of the platform. Any authenticated tenant user with the right per-tool permission can call them.

How it composes

• Spawn under a persona. pj_v3_spawn_agent creates an agent instance under a persona template with an explicit authority level and an attached USD budget. The persona — not the calling user — owns the policies the agent inherits.
• Hylomorphism (delegate down, escalate up). pj_v3_delegate_task hands work to a child; pj_v3_complete_task returns results upward; pj_v3_escalate_task kicks decisions above the agent's authority back to the parent.
• Memory survives instance death. pj_save_memory / pj_get_memories persist what an agent learns under the persona record, so the next instance in that role inherits it.
• Budgets are enforced at dispatch. pj_set_budget / pj_check_budget / pj_get_instance_cost attach a USD cap to a mission or agent; the dispatcher refuses tool calls that would exceed it.
• Audit chain is hash-linked. pj_get_audit_entries returns ordered, SHA-256 hash-chained governance rows. Tamper with a row and the chain breaks at that row.
• Blocking message passing. pj_v3_send_message / pj_wait_for_messages let agents coordinate without polling — the listener blocks until a message arrives.
• Human-in-the-loop. pj_request_approval pauses an agent pending a human decision; pj_approve_decision / pj_reject_decision unblock it. The whole exchange appears in the audit chain.
• Pause, resume, hand off. pj_v3_pause_session / pj_v3_resume_session halt and resume an agent in place; pj_prepare_handoff / pj_resume_from_handoff move a long-running mission across instances without losing context.

Browse the full 9-tool surface at /reference/tools/pjv3/. Walk through a first spawn at /recipes/spawn-agent/.

Production evidence — an accounts-payable validation mission

Parent agent intakes invoices, child validator agent verifies them, escalates above its authority where required, parent approves, child retires. This mission was orchestrated through the PJ Agent Orchestration System — pj_v3_spawn_agent for the validator, pj_check_budget for the cap, pj_get_audit_entries for the chain.

• Mission cost: $6.13 against a $10 hard budget — 61% used.
• Cache hit rate: 94.3% across 9 M cache-read tokens.
• Spawn-to-retire wall-clock: ~10 minutes, end to end.
• Audit chain: 9 unbroken hash-linked rows for the parent.

Costs are reported by the underlying model provider; cache-hit rate and audit chain integrity are measured in-platform.

Same permissions, same data, same observability

• Same permissions. Whether a call comes from a PJ-orchestrated agent or a human, it hits the same dispatcher and the same RBAC check. Per-org scope catalogs at /.well-known/oauth-authorization-server/groups/{org}/ apply identically.
• Same data. Agents read and write the multi-tenant business database — the same one that powers the 17 applications. No separate "agent data plane".
• Same observability. Per-agent cost telemetry rolls into the mission view; pj_get_audit_entries returns ordered, hash-verified rows for any mission.

Next steps

• 3-step MCP quickstart — connect any MCP-aware client over OAuth 2.1 + PKCE.
• Spawn your first agent — end-to-end walkthrough using the PJ Agent Orchestration System.
• PJ Agent Orchestration System reference (9 tools) — browse the full surface in the handler catalog.