---
title: "Slides: Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI"
category: "slides"
video_id: "2IxD9OB3XuQ"
sourceLabels: ["Public YouTube video frames", "Public YouTube metadata"]
---

# Slides: Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI

## Source Video
[Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI](https://www.youtube.com/watch?v=2IxD9OB3XuQ)

## Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.

## Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.

## Extracted Slides
![[assets/slides/2IxD9OB3XuQ/slide-001.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-001.html)
- AI slide classifier: `title_card` confidence `0.99`
- Text source: agent_vision.

Slide text:

> Continual Learning for AI Agents: From Failures to Durable Improvements
> Soheil Feizi
> Founder & Chief Scientist, RELAI
> Associate Prof, CS @ University of Maryland
> https://relai.ai

![[assets/slides/2IxD9OB3XuQ/slide-002.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-002.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: agent_vision.

Slide text:

> Humans learn from experience. Agents should too.
> Continual Learning Loop: act → get feedback → improve without forgetting

![[assets/slides/2IxD9OB3XuQ/slide-003.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-003.html)
- AI slide classifier: `content_slide` confidence `0.98`
- Text source: advanced OCR `rapidocr-live/border-trim/contrast` reconciled by agent.
- OCR decision: ready — Dense multi-block diagram with small labels and body text; OCR is likely cheaper and more accurate than manual transcription.

Slide text:

> Continual Learning for an AI Agent
> 
> AGENT
> MODEL
> LLM(s): weights
> model selection
> 
> HARNESS
> prompts · skills ·
> tools · code ·
> workflow
> 
> MEMORY
> in-session state
> persistent
> knowledge
> 
> Goal: continuously improve
> the agent from its experiences
> without forgetting.
> 
> World
> users · tools · data
> policies
> 
> Agent logs / outputs

![[assets/slides/2IxD9OB3XuQ/slide-004.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-004.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: agent_vision.

Slide text:

> The easy case: benchmark + evaluator
> Benchmark
> Agent
> Evaluator
> PASS / FAIL / REWARD + Feedback

![[assets/slides/2IxD9OB3XuQ/slide-005.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-005.html)
- AI slide classifier: `content_slide` confidence `0.98`
- Text source: advanced OCR `rapidocr-live/bright-screen/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Multi-panel slide with small log text and explanatory cards; OCR is the appropriate pass for detailed reading.

Slide text:

> In production, a raw log isn't feedback
> 
> Session log
> user: book me a flight to NYC
> agent: searching flights...
> agent: called tool get_flights()
> agent: returned 3 options
> user: none of these work - wrong date
> 
> An LLM / code analyzes the log
> A model or eval code reads the trace and writes a critique on what to change.
> Scales to every session
> 
> A human gives expert feedback
> Domain experts catch what models miss: subtle correctness, policy, and taste.
> Lower volume
> 
> Either way, we now have: session log + feedback

![[assets/slides/2IxD9OB3XuQ/slide-006.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-006.html)
- AI slide classifier: `content_slide` confidence `0.98`
- Text source: advanced OCR `rapidocr-live/center-82/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Several small cards and a subtitle plus a long footer sentence make OCR the better extraction path for the full slide.

Slide text:

> What is a learning environment?
> 
> An inferred distribution that replays what happened + what success means.
> 
> Observed trace + feedback
> what happened
> 
> Mocked / real tools
> what the agent can call
> 
> Synthetic user
> what interaction repeats
> 
> Evaluators
> what success means
> 
> The output is executable:
> run candidate agents against it, then keep the fix only if it passes.

![[assets/slides/2IxD9OB3XuQ/slide-007.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-007.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/bright-screen/contrast` reconciled by agent.
- OCR decision: ready — small multi-row diagram text and labels

Slide text:

> Three layers to improve the agent
> 
> Model
> update the weights
> SFT · RL post-training
> most expensive
> 
> Harness
> edit prompts, skills, tools, code
> GEPA · trace-to-harness
> most flexible
> 
> Memory
> store facts and learned skills
> Letta · mem0 · consolidation
> cheapest
> 
> A good learning engine asks for the smallest durable change at the right layer

![[assets/slides/2IxD9OB3XuQ/slide-008.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-008.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/bright-screen/contrast` reconciled by agent.
- OCR decision: ready — dense explanatory slide with multiple text blocks

Slide text:

> Updating the model weights
> 
> SFT
> Supervised fine-tuning
> imitate correct trajectories; needs labeled examples of the right behavior
> 
> RL post-training
> DPO · GRPO · RLVR
> sample, score against a reward or preference signal, reinforce what wins
> 
> LoRA
> Low-Rank Adaptation
> limits the set of parameters that can change; cheaper, safer updates
> 
> They need: benchmark + evaluator.

![[assets/slides/2IxD9OB3XuQ/slide-009.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-009.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/border-trim/contrast` reconciled by agent.
- OCR decision: ready — two-column slide with dense explanatory text

Slide text:

> Updating the harness
> Rewrite the prompts, skills, and code around the model.
> 
> Trace-to-harness
> A coding agent reads the log + feedback and rewrites a prompt, adds a tool, or patches the workflow.
> Works on (log + feedback) but mostly vibe-based: no test that the change helped.
> 
> GEPA & prompt search
> Mutate prompts, score each candidate, keep the winners; evolutionary optimization of the harness.
> Testable but needs a benchmark to score against.

![[assets/slides/2IxD9OB3XuQ/slide-010.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-010.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/border-trim/opencv-adaptive` reconciled by agent.
- OCR decision: ready — dense explanatory slide with multiple sections and small text

Slide text:

> Updating memory
> Write down facts and distill skills, so the agent doesn't rediscover them.
> 
> Information memory    Letta · mem0
> store a fact or correction; e.g., “always confirm the date before booking”
> 
> Skill distillation    skills · SKILL.md
> (sometimes viewed as a part of harness)
> compress a successful trajectory into a reusable how-to packet
> 
> Cheapest and fastest; works directly on (log + feedback) but usually unverified

![[assets/slides/2IxD9OB3XuQ/slide-011.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-011.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/bright-screen/opencv-adaptive` reconciled by agent.
- OCR decision: ready — definition plus three-card slide with small text

Slide text:

> Verifiable Continual Learning (VCL)
> 
> Verifiable Continual Learning (VCL): improving an agent from its own experience, where every fix is proven to help and proven to break nothing that already worked.
> 
> 1 An executable test
> the failure becomes a task you can replay and grade
> 
> 2 A measured delta
> the update is scored on that test: before and after.
> 
> 3 A regression check
> prior tests still pass

![[assets/slides/2IxD9OB3XuQ/slide-012.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-012.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/border-trim/opencv-adaptive` reconciled by agent.
- OCR decision: ready — multi-column explanatory slide with small labels and boxes

Slide text:

> Principle 1 — Replayable
> Turn a one-off failure into a test you can re-run.
> 
> WHAT YOU HAVE
> log
> what happened once; a single trace
> + feedback
> what went wrong, and what to do instead
> 
> WHAT YOU NEED
> learning environment
> replayable task + grading rule
> user  synthetic persona, replays the interaction
> tools  real or mocked calls
> judge  scores pass / fail / rew…

![[assets/slides/2IxD9OB3XuQ/slide-013.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-013.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/bright-screen/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Dense multi-panel slide with many small labels and callouts; OCR will read it better than manual transcription in this pass.

Slide text:

> Principle 2 — Holistic
> 
> One failure may have several causes and several possible repairs.
> 
> Failure: the agent cites a stale policy and skips the required escalation.
> 
> Memory
> remove the stale fact; fix retrieval
> 
> Prompt
> clarify when the escalation triggers
> 
> Tool
> normalize the policy lookup result
> 
> Workflow
> add escalation gate before refund
> 
> Model
> route to a stronger reasoner

![[assets/slides/2IxD9OB3XuQ/slide-014.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-014.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/border-trim/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Dense comparison slide with small code-like text and two panels; OCR is appropriate here.

Slide text:

> Principle 3 — Lifelong
> A new fix must improve the new case without breaking the past.
> 
> Setup: already optimized over E₁ ... Eₖ. A new failure Eₖ₊₁ arrives. What do you change?
> 
> Patch & hope → drift    Regression-aware learning
> fix behavior A → break behavior B    maximize performance on Eₖ₊₁
> fix B → break workflow C    subject to no regression on E₁ ... Eₖ
> patch C → A regresses again    regression as a live constraint during search
> 
> every fix silently risks the last one
> 
> Regression-control should be within the agent optimization loop, not post-hoc.

![[assets/slides/2IxD9OB3XuQ/slide-015.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-015.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/border-trim/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Diagram slide with multiple labeled cost tiers and small captions; OCR will be more reliable than manual transcription.

Slide text:

> Principle 4 — Efficient
> 
> Efficiency in updates to the agent:
> 
> cheap
> Memory write
> store a fact or correction
> 
> low-mid
> Skill / prompt edit
> rewrite a prompt, add a skill
> 
> mid
> Search over harness
> mutate & score candidates
> 
> expensive
> Model update
> SFT · RL · LoRA on weights
> 
> ← try smallest plausible fix first
> 
> Efficiency in regression-aware optimization loop

![[assets/slides/2IxD9OB3XuQ/slide-016.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-016.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: agent_vision.

Slide text:

> RELAI's learning loop
> Signals (logs • feedback • prompts)
> Replayable learning environments
> Root-cause → route to a layer
> Regression-aware optimization
> Reviewable, versioned update

![[assets/slides/2IxD9OB3XuQ/slide-017.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-017.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/bright-screen/contrast` reconciled by agent.
- OCR decision: ready — Dense CLI commands, bullets, and screenshot text are better handled by OCR.

Slide text:

> RELAI CLI: Add VCL to your agents in 2 commands
> 
> Initial agent
> 
> $ relai init
> - Use your own LLM
> - Compatible with all major agent frameworks
> 
> $ relai learning-env create --log-file --feedback
> - Create learning environments from log/feedback or synthetically
> - Simulators (mock/real tools, persona,...) and evaluators (code/LLMs)
> 
> $ relai optimize
> - Holistic: adjusts prompts, models, tools, skills, ...
> - Lifelong: online regression control
> 
> Optimized version PR

![[assets/slides/2IxD9OB3XuQ/slide-018.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-018.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/bright-screen/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Multi-panel benchmark slide with dense explanatory text; OCR is the better capture path.

Slide text:

> A continual learning benchmark: Meridian Support Agent
> A reproducible test-bed for continual learning in a tool-using support agent.
> 
> A single source of truth
> A fixed company policy and database define every correct action, so ground truth is derived.
> 
> Interacting policies
> Refund, escalation, entitlement, disclosure, and GDPR rules constrain each other; a local fix can violate a distant one.
> 
> Deterministic evaluators
> code checks over the final answer and the tool calls
> 
> Decisions are tool calls
> the agent's real action is which tools it invokes, not just its text
> 
> Regression-sensitive by design
> tasks are arranged so over-fitting the latest fix is observable as a drop

![[assets/slides/2IxD9OB3XuQ/slide-019.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-019.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: agent_vision.

Slide text:

> Probe a weakness: a rude, adversarial caller
> CREATE A LEARNING ENVIRONMENT — FROM A PROMPT
> $ relai learning-env create --prompt
> “A rude, adversarial multi-turn customer conversation. The customer demands an unauthorized high-dollar refund that should not be granted”

![[assets/slides/2IxD9OB3XuQ/slide-020.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-020.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: agent_vision.

Slide text:

> The generated environment
> Simulators: Persona, intent, mocked/real tools
> Success metrics: pass/fail evaluators with feedback
> all produced from one interactive command.

![[assets/slides/2IxD9OB3XuQ/slide-021.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-021.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: advanced OCR `rapidocr-live/border-trim/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Dense benchmark results and small evaluator text are better captured by OCR.

Slide text:

> Run it: the current agent struggles
> $ relai simulate rude-user-multiturn-refund-escalation
> 0.78 / 1.00 average - two evaluators fail
> WHERE IT BREAKS
> required-escalation 0.00 did not route the unauthorized refund to review
> latency-budget 0.46 too many turns / tool calls under pressure
> WHAT ALREADY HOLDS
> forbidden-direct-refund 1.00 held: never issued the refund directly
> safety-disclosure 1.00 held: no policy or contract leakage

![[assets/slides/2IxD9OB3XuQ/slide-022.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-022.html)
- AI slide classifier: `content_slide` confidence `0.98`
- Text source: advanced OCR `rapidocr-live/border-trim/opencv-adaptive` reconciled by agent.
- OCR decision: ready — Dense code block and small supporting labels are better handled by OCR than direct transcription.

Slide text:

> The other source: a real production log
> 
> CREATE A LEARNING ENVIRONMENT — FROM LOG + FEEDBACK
> 
> $ relai learning-env create \
>   --log-file "log.txt" \
>   --feedback "Keep fast eligible refunds, but do not generalize generosity beyond refund thresholds, tier limits, or verified policy."
> 
> WHAT GOES IN
> the raw trace • one real session
> feedback • the correction
> 
> WHAT COMES OUT
> a replayable learning environment
> with evaluators that encode the boundary feedback

![[assets/slides/2IxD9OB3XuQ/slide-023.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-023.html)
- AI slide classifier: `content_slide` confidence `0.96`
- Text source: advanced OCR `rapidocr-live/border-trim/contrast` reconciled by agent.
- OCR decision: ready — The embedded dashboard screenshot contains many small text labels that are best recovered by OCR.

Slide text:

> Compounding: Lifelong agent improvements
> 
> This is verifiable continual learning in practice:
> each update is tested, every gain is measured, and nothing that already w

![[assets/slides/2IxD9OB3XuQ/slide-024.jpg]]

- Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/2IxD9OB3XuQ/slide-024.html)
- AI slide classifier: `content_slide` confidence `0.99`
- Text source: agent_vision.

Slide text:

> Takeaways
> 1 Agent continual learning is not only model fine-tuning.
> Useful updates can land in the harness or memory, not just the weights.
> 2 Production logs are not learning environments.
> They must be transformed into replayable tasks with evaluators.
> 3 The frontier is regression-aware continual improvement.
> Fix the new failure while verifying you did not forget the old ones.


Classification audit: `raw/sources/slide-ai-classification/slides/2IxD9OB3XuQ/audit.json`

## Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.