Dynamic Agent Specialisation by Joel Jeon
Describe the job. DAS constructs complete specialists and switches only when a challenger proves better.
Dynamic Agent Specialisation is a new technological advancement that compiles a bounded business role into several complete AI specialists, tests their actual external outcomes against strong baselines, and recommends a switch only when a challenger proves it should.
The control loop is the product. Each part exists to preserve authority, evidence and the original goal.
The input is not just a job title. It includes the desired business outcome, systems, policies, authority, examples, independent success criteria and the company’s quality, cost and speed priorities. An existing agent can enter as the baseline.
Each candidate binds a model, instructions, context, tools, memory policy, allowed authority, escalation behaviour, independent verifier, cost limits and latency limits. The configuration being tested is the whole worker, not only its prompt.
Wrong tools, missing context, mismatched verification, excessive authority, unsafe memory scope, impossible budgets and invalid escalation rules are hard failures. Safety and authority cannot be averaged away by a high performance score.
DAS does not merely rank candidates it generated itself. The bounded campaigns compare them with strong general agents, ordinary manually configured agents and expert configurations where those alternatives can be run fairly.
Candidate and baseline identities are fixed before prospective unseen cases are released. Development, validation, adversarial, unseen and repeatability evidence stay separate so a good-looking result cannot be created through post-hoc score chasing.
A separate external-state verifier checks whether the intended order, ticket, CRM record or other business result exists and whether forbidden side effects occurred. The candidate is not allowed to grade itself.
Missing outcomes, incorrect effects, handoffs, tool use, latency and cost remain in the evidence. A narrow repair may address a missing outcome. Incorrect or unsafe effects stop instead of disappearing behind an automatic retry.
A generated specialist activates only when it wins under the decision rule fixed in advance. Otherwise DAS recommends keeping the existing setup. Retaining the current agent is a valid and sometimes correct output.
Current local adapter work can import an explicitly selected safe subset of OpenAPI operations or a pinned customer-local Model Context Protocol tool list. Exact tool allowlists, authority, uncertain-write reconciliation and external-state visibility remain attached.
A later local model-backed portfolio run completed three new fictional assignments through separate external-state checks, with no observed unsafe attempt in those tasks. No real customer system has been activated.
One unchanged compiler and evaluation mechanism has operated across fictional procurement, software support and customer-revenue operations worlds.
A compiler-built specialist earned selection in one role. The existing specialist stayed in two because the challengers did not prove a better overall choice.
The repository has passed more than 250 automated local checks across policy, persistence, evaluation and evidence layers. These are codebase checks, not 250 model-agent trials.
Agent construction, orchestration and evaluation are active categories. DAS does not claim those components are absent. Its focus is joining them into one evidence-gated compilation and switch-or-retain decision.
Agent evaluation, prompt versioning and optimization.
Official source ↗CrewAIAgents, crews, flows, memory, guardrails and deployment.
Official source ↗Microsoft Copilot StudioAgent construction, tools, evaluation and governance.
Official source ↗DSPyOptimization of instructions and examples inside language-model programs.
Official source ↗DAS joins complete specialist-package construction, hard authority and safety contracts, strong-baseline comparison, independent external-state verification, and a final switch-or-retain decision. As of August 2026, my review of public product documentation has not found another system demonstrating this same specific loop.
The bounded three-role technical mechanism is complete. Customer deployment, production reliability, real-company setup-time savings and demand remain unproved.