WorkingAgents + Adaption Labs: Governed Agents That Learn on the Fly

By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 7, 2026, 21:07

Adaption Labs builds AI that evolves through interaction. WorkingAgents governs AI agents in production. Together, they solve a problem neither addresses alone: how do you let AI models continuously learn from live business data while maintaining the permissions, audit trails, and guardrails that production environments demand?

Continuous learning without governance is a compliance nightmare. Governance without adaptability locks enterprises into static models that drift from reality. The intersection is where enterprise AI actually works.

What Adaption Labs Brings

Adaption Labs was founded by Sara Hooker (former VP of Research at Cohere, Google DeepMind veteran) and Sudip Roy (former Director of Inference Computing at Cohere). They raised a $50M seed round — one of the largest in 2026 — led by Emergence Capital, with Mozilla Ventures, Fifty Years, Threshold Ventures, and Alpha Intelligence Capital participating. Valuation: $1B.

Their thesis: the days of monolithic AI are over. The future belongs to smaller, efficient models that adapt continuously rather than massive models retrained in expensive cycles.

Core Technology — Gradient-Free Learning:

The fundamental innovation. Traditional AI training adjusts billions of internal weights through computationally expensive gradient descent (backpropagation). Adaption modifies model behavior at inference time — when the model is actually responding to queries — leaving core weights untouched. As co-founder Sudip Roy explains: “Our models learn from that interaction immediately. If you tell it ‘that’s wrong, use this policy instead,’ it adapts its weights on the fly, for that specific context, with negligible compute overhead.”

Key techniques:

Three Product Pillars:

  1. Adaptive Data (in early access) — Platform for dynamically shaped datasets at scale. 82% average increase in data quality across early deployments. Supports 242 languages. Focuses on the rare, nuanced edge cases where static datasets fail — the “long tail” where real business value lives.

  2. Adaptive Intelligence (waitlist) — Next-generation AI evolved for any industry, language, or specialization. The vision: deploy a smaller, cheaper base model that “grows into” its role through real-world interaction. A legal AI starts with general knowledge and, through corrections by senior partners, evolves into a specialized expert — no GPU-intensive training required.

  3. Adaptive Interfaces (waitlist) — Innovation hub reimagining human-AI interaction beyond the standard chat bar.

Enterprise Value Proposition:

What WorkingAgents Brings

WorkingAgents is an AI governance platform — the access control layer and orchestration engine between AI agents and the internal and third-party systems they interact with. Three gateways, one control plane:

Capability-based access control means each agent sees only the tools it’s authorized to use. Virtual MCP Servers define permission boundaries per team, per role, per use case — configuration, not code. Every action is logged with full context: who triggered it, what tool was called, what guardrails fired, what it cost.

What WorkingAgents doesn’t have: adaptive AI models, continuous learning infrastructure, or gradient-free inference optimization. What it does have: the governance framework that makes continuous learning safe for production.

The Strategic Thesis

Adaption’s gradient-free learning creates a new category of risk that traditional AI governance wasn’t designed for.

Static models are governable because they’re predictable. You test a model, validate its outputs, deploy it, and it behaves the same way tomorrow as it does today. Governance is a point-in-time certification.

Adaptive models break that assumption. A model that learns on the fly — that changes its behavior based on interactions — requires governance that’s equally dynamic:

WorkingAgents provides exactly this infrastructure: permissions on who can trigger what, guardrails on acceptable boundaries, and audit trails on everything that happens.

The Gap Analysis

Adaption Labs Gap WorkingAgents Solution
No permission model for who can trigger model adaptation Capability-based access control — define who can teach, correct, or reshape models
No audit trail for what the model learned and when Every interaction logged with full context, cost, and guardrail evaluation
No guardrails on what the model can learn from Pre/during/post execution safety checks on adaptation triggers
No multi-agent orchestration for adaptive workflows Agent Gateway manages complex, multi-step workflows with retries and human-in-the-loop
No human-in-the-loop approval for high-stakes adaptations Real-time approval workflows: “Model wants to update its contract interpretation — approve?”
No cross-system deployment governance Virtual MCP Servers scope different adaptation policies per team/role/department
No governed access to enterprise tools and data sources MCP Gateway orchestrates access to internal and third-party systems with least-privilege permissions
WorkingAgents Gap Adaption Labs Solution
Static model routing — no continuous learning Models that adapt at inference time with negligible compute overhead
LLM routing optimizes for cost/latency, not task fit Adaptive Intelligence matches computational spend to problem difficulty
No domain specialization for governed models Base models evolve into domain experts through interaction
Data quality for enterprise knowledge is manually curated Adaptive Data delivers 82% quality improvement across 242 languages
No edge/lightweight deployment option for AI models Gradient-free adaptation runs on edge hardware, not just H100 clusters
Agent responses don’t improve from user corrections On-the-fly learning incorporates corrections immediately

Synergy Areas

1. Governed Continuous Learning

The highest-value integration. Adaption’s models learn from every interaction. WorkingAgents governs which interactions are allowed to teach:

Adaptation Governance Policy — Legal Department

Who Can Teach:
  ✓ Managing Partner — full adaptation authority (contract law, liability, compliance)
  ✓ Senior Associate — domain adaptation (their practice area only)
  ✓ Junior Associate — correction flagging only (queued for senior review)
  × Paralegal — read-only, no adaptation triggers
  × External Client — zero adaptation authority

What Can Be Learned:
  ✓ Domain terminology and precedent interpretation
  ✓ Firm-specific formatting and citation preferences
  ✓ Client-specific communication tone
  × Bypass of ethical guidelines
  × Modification of compliance rules
  × Alteration of conflict-of-interest checks

Audit Requirements:
  ✓ Every adaptation logged with before/after comparison
  ✓ Weekly adaptation report to compliance officer
  ✓ Quarterly adaptation review by ethics committee

WorkingAgents’ Virtual MCP Servers enforce these policies at the gateway level. The governance is configuration, not code. Different departments get different adaptation authorities. The same Adaption model serves the entire firm, but each team’s ability to reshape it is precisely scoped.

2. Governed Access to Enterprise Data for Adaptation

Adaption’s models improve through interaction with real business data. WorkingAgents controls which data sources the adaptive model can access — and what it’s allowed to learn from each:

The adaptive model gets smarter through governed interactions with enterprise systems. WorkingAgents ensures those interactions respect organizational boundaries.

3. Cost-Optimized Agent Routing

WorkingAgents’ AI Gateway currently routes by cost and latency — simple queries to cheaper models, complex reasoning to capable ones. Adaption’s Adaptive Intelligence adds a third dimension: task fit.

Instead of routing every query to a general-purpose LLM:

The result: lower cost (smaller adaptive models vs. large general models), better accuracy (domain-specialized through continuous learning), and full governance (every routing decision logged and auditable).

4. Adaptive Data for Enterprise Knowledge Pipelines

Adaption’s Adaptive Data platform (82% quality improvement, 242 languages) integrates with WorkingAgents’ orchestration layer to improve the quality of data flowing through governed pipelines:

5. Edge-Deployed Governed Agents

Adaption’s gradient-free models run on edge hardware — consumer-grade devices, not H100 clusters. WorkingAgents’ governance extends to these edge deployments:

The combination unlocks AI in environments where cloud-only solutions fail — either because of latency, bandwidth, privacy, or regulatory constraints.

6. Adaptation Rollback and Version Control

When an adaptive model learns something wrong, you need to undo it. WorkingAgents’ audit trail enables governed rollback:

{
  "adaptation_id": "adapt-2026-03-07-1423",
  "model": "legal-contract-review-v3",
  "triggered_by": "senior_associate_chen",
  "adaptation_type": "interpretation_correction",
  "context": "Force majeure clause interpretation for pandemic scenarios",
  "before": "Standard pre-2020 interpretation excluding epidemics",
  "after": "Updated interpretation including pandemic provisions per Chen v. GlobalCorp",
  "approved_by": "managing_partner_rodriguez",
  "status": "active",
  "rollback_available": true
}

If a later review determines the adaptation was incorrect, the managing partner triggers a rollback through WorkingAgents. The governance layer ensures only authorized users can revert adaptations, the rollback itself is logged, and downstream agents that consumed the adapted model are notified.

Partnership Model

Phase 1: Governed Adaptation Prototype (Weeks 1-6)

Phase 2: Adaptive Data Integration (Weeks 7-12)

Phase 3: Enterprise Reference Architecture (Weeks 13-20)

Why This Partnership Works

The AI industry has two camps that rarely talk to each other.

The model camp builds better models — more capable, more efficient, more adaptive. Adaption Labs is at the frontier, with gradient-free learning that lets models evolve through interaction without expensive retraining. Sara Hooker’s thesis that smaller, adaptive models beat larger static ones is gaining traction as enterprises hit the cost wall of scaling.

The governance camp builds control planes — permissions, audit trails, guardrails, compliance. WorkingAgents sits here, providing the access control layer and orchestration engine that makes AI agents accountable across every tool call and every decision.

But adaptive AI makes them inseparable. A model that learns on the fly is a model that changes in production. A model that changes in production without governance is a liability. A governance system that doesn’t understand adaptation is governing yesterday’s model.

Adaption makes AI that evolves. WorkingAgents makes AI that’s accountable. The partnership makes AI that evolves accountably — smaller models that grow into domain experts through real business interactions, governed at every step, auditable at every adaptation, and reversible when things go wrong.

The days of monolithic AI may be over. The days of ungoverned adaptive AI should never begin.

James Aspinwall is the founder of WorkingAgents, an AI governance platform specializing in agent access control, security, and integration services for enterprises deploying AI at scale.