AI Agents at Work: session notes

AItalks2Grow, Warsaw, 18 October 2026. Speaker: Andrew Tomin. English.

18 October 2026 Warsaw English No coding required
About these notes. Compiled from the session transcript and checked against public sources, linked inline. Where something said on stage was imprecise, the text says "Correction". Prices, plan limits and product behaviour change quickly, so check current details before you decide. Disclosure: the Wispr Flow link is a referral link, so Andrew may receive a reward if you sign up. The OpenChamber link carries a tracking tag with his name.

Who this is for and how to read it

For product managers, business analysts, project and team leads who do not code. If you already know the difference between a chat and an agent, skip to Part 3.

Five things to take away AItalks2Grow · if you read nothing else 1 An agent works in a loop on your files A chat answers. An agent reads, acts, checks and repeats, and you steer. 2 A skill is a reusable instruction package Write it with AI, then read and edit it yourself before you trust it. 3 Outputs vary from run to run Same input, different result. Your job moves from writing to checking. 4 Check in layers A different model, then tests or a checklist, then the person who owns the task. 5 Start with one task and a small pilot Measure time and rework before and after. Find the people already experimenting.
Swipe sideways to see the full diagram

Words used in these notes, in plain language

WordWhat it meansEveryday comparison
AgentAn AI that does the work on your files and tools, not only answers questionsA colleague who opens the files and does the task
SkillA reusable set of written instructions for one kind of taskThe picture steps in flat-pack furniture
ContextEverything the agent knows about the task: your files, rules and questionThe briefing you give before work starts
Context windowHow much the agent can keep in mind at one timeThe size of a desk: it fills up
ConnectorA link that lets the agent reach Jira, Drive, Figma and similar toolsA plug into another system
Sub-agentA helper with its own context for one small jobAn assistant with their own desk
Handoff fileA saved note that tells the next session or tool where things standA note left for the next shift
SpecA short document that says what to build and how you will check itA brief with a checklist
PrototypeA clickable draft, not real softwareA mock-up you can walk through
CLAUDE.md, AGENTS.mdThe rules file a tool reads first in your project folderThe house rules on the door

Part 1. Chat versus agent

In plain words: a chat is a conversation. An agent is a colleague who can open your files, do the task and show you the result to check.
Chat vs agent AItalks2Grow · same model, different way of working Chat: talk Agent: do the work You ask, it answers You carry the context: paste, upload, re-explain You copy the result into your own files Connectors and file tools may exist, but you steer each step You give a goal and rules (context, skills) It reads and writes files in your project folder It runs tools, checks its own step and repeats You review the result and decide The difference is not the model. It is the loop on your files, and who carries the context.
Swipe sideways to see the full diagram
  • In a chat you talk. An agent does the work: it reads your files, runs tools, checks its step and saves results.
  • Correction from the session: chat products today also have connectors, skills and file creation (Anthropic, 16 September 2026). The practical difference is local file and tool access plus a loop where the agent acts and checks, with you steering less step by step.
  • An agent is more than a model. It has context (the files and instructions you give it), skills (reusable instructions), memory, tools (build a doc, build an HTML page, analyse a CSV) and connectors (Jira, Confluence, Google Drive, Miro, Figma, repositories).
What an agent is made of AItalks2Grow · model in a loop with context, skills, memory and tools Context files, instructions, the task Skills reusable how-to (SKILL.md) Memory what to keep between sessions Model reads, decides the next step Tools build a doc or HTML, analyse a CSV Connectors Jira, Confluence, Drive, Miro, Figma Sub-agents own context window per sub-task Your computer project folder, browser, terminal Agent loop Understand Plan Do Check Repeat
Swipe sideways to see the full diagram
  • Examples of agent tools named in the session: Claude Code, Codex, Google Antigravity. Andrew mostly works in VS Code with the Claude Code and Codex extensions.

Part 2. The workflow shown

In plain words: we gave the AI a messy meeting record, answered its questions, and got a clickable draft of a new feature. Then we checked the draft against what we had agreed.
From meeting to verified outcome AItalks2Grow · the workflow shown on stage, with Human Check at the gate INPUT Raw meeting transcript Delivery problem: recipient is not home SYSTEM meeting-to-spec skill Questions to you, answers, then a lightweight spec HUMAN CHECK You read the spec Proceed, revise or pause OUTPUT Prototype Handoff Claude to Codex, clickable HTML, fake data only VERIFIED Check vs the spec Different model or a person compares result with each check Tools in this run: Claude Code (spec + handoff) | Codex (build) | fresh Claude Code (validate) What happened live Spec and prototype were produced and clicked through. The last step (validate-outcome) failed on stage because of a setup problem. Outputs differ each run, so every step needs a person to look at the result. Input, system, output, verified: the same four states behind every AI Sandbox Lab case.
Swipe sideways to see the full diagram

The input was a meeting transcript about one problem: a delivery recipient cannot be home, and wants to reschedule so the courier does not come in vain.

  1. Raw transcript is the starting point. Work conversations are the foundation for tangible output.
  2. A skill called meeting-to-spec turns it into a lightweight spec. The skill contains a spec template and a description of how to do the job.
  3. The agent asks clarifying questions. Better answers give a better result. Answers were given in English.
  4. Human review of the spec. Andrew noted it will not match your own template unless you give the agent your template.
  5. Handoff: Claude wrote a handoff file for Codex, which then built the prototype with a spec-to-prototype skill. The agent first proposed a screen plan and asked whether to proceed.
  6. The result was a clickable HTML prototype with no backend: an SMS arrives, the user picks another delivery slot and confirms.
  7. Verification: Codex wrote a verification handoff, and Claude was asked to validate the outcome against the spec with a validate-outcome skill.

What did not work: the validation step failed in the session ("something went wrong", attributed to setup or timing). Andrew pointed participants to the prompts on his site to run it themselves. The prototype also showed a map that was not needed, and the second run looked different from the first.

Part 3. Key ideas

In plain words: write instructions once, reuse them, and always look at the result before it leaves your desk.

Skills are reusable instruction packages. Andrew's analogy: the assembly instructions that come with IKEA furniture. The parts are your files and tools, the instruction sheet is the skill, and the finished piece is your spec, report or prototype. A good skill has a clear name and clear steps. Technically it is a folder with a SKILL.md file; the agent reads the name and description first and loads the rest only when the task matches (Codex docs). The same format works in Claude Code and Codex.

A skill is like furniture instructions AItalks2Grow · the picture steps that come in the flat-pack box The parts your files, data and tools + The instruction sheet 1 2 3 the skill: steps in a fixed order = The finished piece a spec, report or prototype Without instructions Each try gives a different chair. You re-explain the job every time. With instructions Same steps, a result you can repeat and check against the same list. Read the instructions before you follow them: that is the security check for skills.
Swipe sideways to see the full diagram
Inside a skill AItalks2Grow · a folder the agent reads in stages meeting-to-spec/ SKILL.md name + description, then the steps spec-template.md optional: the template to fill scripts/ optional: code the agent can run Scripts are code. Read them before you trust the skill. 1 You ask for a task 2 Agent reads each skill's name and description 3 It picks the matching skill 4 It loads the full instructions 5 It follows the steps, then you review The same SKILL.md format works in Claude Code and Codex.
Swipe sideways to see the full diagram

Do not hand-write skills; write them with AI and then edit. Claude and Codex have a built-in skill creator. Speak your intent (Andrew uses voice dictation with Wispr Flow), then read and edit the markdown file. With many skills, consider a skill for reviewing and verifying skills.

Skills from outside need a security check. Skills can contain executable code. Anthropic's own guidance is to install only from trusted sources and to audit anything else by reading every bundled file (Anthropic). Correction from the session: asking the agent to review a skill helps, but it is not enough, because a few lines of plain instructions can also be harmful. A person reads it too. One study reported by a news site found about a quarter of 31,132 public skills had at least one vulnerability (AI Weekly); the figure has not been independently verified.

Before you install a skill AItalks2Grow · a skill is instructions, sometimes with scripts the agent can run 1 Source Trusted author or the vendor's own repository? 2 Read it Open the whole folder, not only SKILL.md 3 Look for scripts Any code, network calls, file access outside the project? 4 Try safely Test on fake data in a throwaway folder first Why plain-text skills can also be risky A few lines of instructions can tell an agent to read private files or send data out. Ask an AI to review a skill if you like, but a person reads it before it is trusted. Reported numbers (one study, via news summary) 26.1% of 31,132 skills had at least one vulnerability. Skills with scripts were about 2.1 times more likely. Source: aiweekly.co. Treat as indicative, not exact.
Swipe sideways to see the full diagram

Be specific about output. "Summarise this" gives vague results. Ask for a fixed number of bullets, and for proof of each statement from the source.

Outputs vary. Models are probabilistic, so the same input can give a different result, even at temperature 0 (paper). Work shifts from producing to checking, like reviewing a junior colleague. Andrew noted that some people feel only slightly more productive: they can do more but must verify more. One study of experienced developers found them 19% slower with AI tools while they believed they were faster (summary); that was a different group from PMs and BAs, so treat it as a caution, not a rule.

Who checks the result AItalks2Grow · stronger checks sit higher. Use more than one. Human who owns the task knows the domain, signs off Tests and checklist each acceptance check, pass or fail with evidence Different model less biased toward its own style of output Same model re-reads its work weakest: tends to favour its own output Models are probabilistic: the same input can give a different output, even at temperature 0. Research basis: self-preference bias in LLM judges (arXiv 2410.21819); non-determinism (arXiv 2408.04667).
Swipe sideways to see the full diagram

Always verify. Checking with the same model is the weakest option, because models tend to favour their own output (paper). Prefer a different model, then tests or a checklist, then the person who owns the task. Do this especially for anything going to clients, engineers or leadership. Andrew's own example: a day earlier a wrong QR code was generated by Codex because the agents.md file was not set up the way Claude's setup was. "Never trust AI."

Retro with the model. After a bad output, ask the model to update CLAUDE.md, agents.md or the skill so it does not repeat the mistake.

Give design guidance. Without a design system or examples, output looks generic. If your team has a design system, connect Figma and reuse the tokens. On a ten-year-old project the design system may itself be out of date; ask the designer for the latest.

Handoff files are useful when a context window runs out or when switching tools or agents: the next session reuses the artifact instead of losing it in the conversation.

Part 4. Choosing tools and models

In plain words: use a stronger model to plan and a faster one to build, and keep notes in files so nothing is lost when a chat ends.
The setup shown in the session AItalks2Grow · one tool per job. Prices checked on vendor pages, 18 Oct 2026 PLAN AND DOCUMENT Claude $20 plan BUILD FROM A SPEC Codex $100 ChatGPT Pro plan, Codex included DEEP RESEARCH Gemini about $20 a month (Google AI Pro $19.99) W DICTATE PROMPTS Wispr Flow free plan; Pro $15 a month, $12 if billed yearly O ONE WINDOW, SEVERAL MODELS OpenChamber open source, with a context panel READ AND EDIT FILES VS Code Claude Code and Codex extensions Plans and limits change often. Start with any tool, or the one your team already has.
Swipe sideways to see the full diagram
Who does what in a multi-tool setup AItalks2Grow · Andrew's setup, shown as a way to think, not a recommendation to buy Planner (stronger model) spec, architecture, documentation, hard decisions Builder (cheaper, fast) builds from an approved spec or plan (Codex in this demo) Sub-agents research, checks, each in its own context window Glue: handoff files in the project folder. Saving a handoff does not start the next tool; you do. Notes Claude Code has an opusplan option: Opus plans, Sonnet executes. Plan names and prices change often. Check current limits before choosing. Start with any tool. Use what your team already has.
Swipe sideways to see the full diagram
  • Andrew's personal setup: a $20 Claude plan for planning and documentation, a $100 plan for Codex for building from a good spec, a Gemini plan at about $20 a month for deep research (Google AI Pro is listed at $19.99). Correction: the $100 option is a ChatGPT Pro plan that includes Codex (TechCrunch), and Claude Max 5x is also $100, so price tiers match and what differs is usage limits, which change often. Check current plans.
  • Mental model: a stronger model is a senior architect who plans; a cheaper model is a mid-level engineer who builds to the plan; sub-agents are for research and keep your main context clean. This matches Claude Code's opusplan option (docs) and how sub-agents work (Anthropic).
  • Voice dictation: Andrew dictates prompts and edits with Wispr Flow on its paid plan. Flow has a free plan; Flow Pro is $15 a month, or $12 a month billed yearly (Wispr pricing).
  • Choice is partly taste. Some artifacts he likes better from Claude.
  • Advice for beginners: start with anything, do not overthink the model. If you have a work subscription, use what your engineers use.
  • Spec frameworks: Andrew uses OpenSpec at work and Superpowers for private projects; tried BMAD and found it did not fit. Stick to what your team uses.
  • A tool to try: OpenChamber, an open-source desktop, web, mobile and VS Code interface built around OpenCode, with a session context panel. Andrew uses it to work with several models in one place. Its site does not mention Codex, so check which providers you can connect.
  • Claude Cowork versus Claude Code: Andrew reported that Cowork does not show context or token limits, compressed a long chat without warning and lost artifacts. Another attendee lost Cowork sessions after reinstalling the desktop app. Similar user reports exist (issue tracker); these are user reports, not vendor statements, and the product may have changed. Both tools read the same CLAUDE.md in a shared folder (guide). Anthropic's merge announcement says nothing about context or token visibility, so those points remain user reports. Lesson: keep outputs and handoffs as files in your project folder, not in chat history.
  • Update: as mentioned in the room, Anthropic's announcement of 16 September 2026 merges Cowork and chat into one conversation, and Claude Design now also works inside conversations. Claude Code remains a separate tool (announcement, help article). It is rolling out first to Pro and Max plans. The prototype-to-code handoff mentioned in the room is Claude Design sending a bundle to Claude Code (tutorial).
  • Working from a phone: Claude Code has Remote Control to continue a local session from your phone (Help Net Security). Andrew chooses not to use it, to keep a break from the computer.
  • Instruction files: Codex reads AGENTS.md and Claude Code reads CLAUDE.md (guide), which explains the QR bug. Keep one shared set of rules or link them.
Which tool reads which file AItalks2Grow · the tools overlap, so name them carefully Claude chat + Cowork One conversation since 16 Sept 2026 Rolling out to Pro and Max Claude Design Also works inside conversations Hands a bundle to Claude Code Claude Code Separate coding tool Not part of the merge Codex OpenAI's coding agent Skills in the same SKILL.md format Shared project folder CLAUDE.md Read by Claude Code and Cowork AGENTS.md Read by Codex Keep one shared set of rules, or link the two files. Otherwise each tool follows different instructions. Sources: Anthropic announcement and help article, 16 Sept 2026; Codex and Claude Code docs. Logos belong to their owners.
Swipe sideways to see the full diagram

Part 5. Adoption in teams

In plain words: try it on one small task with a few people, and measure the time saved before you roll it out wider.
Start small: a pilot you can measure AItalks2Grow · Q&A answer, turned into steps 1 One task Repeatable, clear input and output 2 Baseline Time and rework today, before AI 3 Small group About 10 people, find the champions 4 Run 2 to 4 weeks Same task, with agent and checks 5 Compare Cycle time, rework, does use spread? Managers who try the tools themselves build better cases Leading by example shapes how fast a team adopts. A vague order to use AI is not a specific use case. Honest limits A full ROI study is a bigger project, but a before/after pilot on one task is cheap. Usage counts show activity, not value. Measure the work that changed.
Swipe sideways to see the full diagram
  • Share practices, not just personal tricks. If a team used to write Jira stories in a format, create a skill for that format and share it.
  • Andrew's observation: where managers use AI coding assistants themselves, their teams adopt faster, because they lead by example.
  • Andrew's observation: some teams already have a product manager plus a tech lead shipping features, and he expects this to become common in a year or two. Quality assurance for fully AI-built work was described as an open problem: it needs the engineering lead (tests, evals, tools such as Playwright were mentioned by an attendee).
  • Benefit shown for non-coders: prototype yourself, validate with customers, then bring a spec and prototype to engineering rather than wait two to three months for capacity.
  • ROI: Andrew's position is that a full ROI study is a large consulting project. A small before/after pilot on one task (time and rework) is far cheaper, and practitioner guidance recommends measuring cycle time, rework and whether use spreads beyond early enthusiasts (guide).
  • Learning to use these tools: Andrew's experience is one and a half to two and a half weeks to start using them at work. He has trained more than 70 non-coders.

Updated since the session

Updated since the session AItalks2Grow · what was said, and what the sources say SAID IN THE SESSION CHECKED AGAINST SOURCES A chat cannot use tools or your computer Chats now have connectors and files. The real difference is the loop on your files. Use approved skills, or ask AI to review Trusted source, plus a person reads every file. AI review alone is not enough. $100 for Codex; Claude costs more The $100 plan is ChatGPT Pro with Codex. Claude Max 5x is also $100. Verify with the model itself or another Same model is weakest. Use a different model, tests, then a person. ROI cannot be calculated cheaply A small before/after pilot on one task is cheap.
Swipe sideways to see the full diagram

Part 6. Appendix: questions from the room

QuestionAnswer given
Did you create the skill yourself?With AI help, then edited and reviewed. Use the built-in skill creator.
Where can we download skills safely?Prefer trusted sources such as the tool vendor's own repositories, and read the whole skill yourself before installing. An AI review alone is not enough (see Part 3).
Why not one Codex subscription instead of Claude plus Codex?Cost and taste. Andrew likes Claude's output for some artifacts. See Part 4 for how the plans compare.
Which spec frameworks have you tried?See Part 4. Depends on what the team uses.
Same data, different prototype?Yes. Outputs vary; guide with samples and a design system.
Do you use Claude Cowork for prototypes?Not much; goes straight to Claude Code. Uses Cowork for office documents.
Can you work from a phone?Andrew deliberately does not. Others use scheduled tasks in Cowork or voice notes to a chat assistant. Remote Control for Claude Code exists.
Workshop resources

Prompt pack and starter files

Repeat the workshop exercise: Claude Code -> Human Check -> Codex -> fresh Claude Code review. The task is: A delivery recipient cannot be home at the assigned time and wants to choose another available slot without calling support.

The transcript, names, dates, delivery slots, call volumes, and costs are fictional. The prototype uses sample data and does not connect to dispatch, send messages, change a real delivery, or prove fewer support calls.

Start with the complete starter

Unzip the complete starter and open the extracted meeting_to_prototype_starter folder in both Claude Code and Codex. Keep the hidden .agents and .claude folders. Start at step 2 below: all three skills are included. The ZIP also contains the fictional transcript, workflow, brief template, underlying prompt sequence and Human Check checklist.

For the optional skill-creation exercise, use the live-build ZIP and start at step 1. Use one folder for the whole run. You switch tools and start each stage; shared files carry the handoff. Read the skill instructions before running them. Review the spec yourself before approving a build.

Fictional clarification answers for this exercise

Supply these after the agent asks its clarification questions. These are exercise choices, not facts from the transcript. If you choose different answers, update the spec and its checks together.

  • The user is a home delivery recipient. Open the sample delivery directly; no SMS screen.
  • Current slot: Thursday 22 October 2026, 09:00-12:00. Alternatives: Friday 23 October, 09:00-12:00 available; Friday 23 October, 13:00-16:00 full; Saturday 24 October, 09:00-12:00 available.
  • Allow one selection. Continue stays disabled until an available slot is selected. Show original and selected slots before explicit confirmation.
  • Cancellation keeps the original slot and must not show a changed delivery.
  • Confirmation shows the chosen slot and "Simulation only. No real delivery has changed."
  • Use English and one offline HTML file. Exclude maps, login, address changes, real connections, contact options and an empty-state scenario. Real availability and cutoff rules remain unresolved.

Step 1

Create the first skill

Optional: skip this step with the complete starter. To create the skill yourself, use the live-build ZIP. It deliberately omits meeting-to-spec so this is a real creation step.

Create a project skill named meeting-to-spec for this workshop.
This is a draft for a three-minute live exercise. Create the files directly;
we will inspect and test the draft in the next step of the demonstration.
Its input is a meeting transcript and the task we want to explore.
It extracts final decisions, constraints and gaps with source references,
asks the human questions that change the prototype, and waits for answers.
It then writes a short spec using handouts/working_brief_template.md.
Keep facts, human answers and assumptions distinct. Preserve English.
The spec needs a version, one user journey, exclusions and observable checks.
Stop for a human's Proceed, Revise or Pause on that version before any build.
Do not invent answers or approval, and do not build the prototype in this skill.
Create the skill in .agents/skills/meeting-to-spec for Codex and expose
the same instructions in .claude/skills/meeting-to-spec for Claude Code.
Put a copy of the template inside the skill so it can be reused elsewhere.
Step 2

Claude Code prepares the brief

Run this in Claude Code after creating or opening the prepared meeting-to-spec skill.

Read .claude/skills/meeting-to-spec/SKILL.md and use that skill on
inputs/meeting_transcript.md. The task is to let a delivery recipient choose
another available delivery slot without calling support. Work in English.
Save outputs in outputs/. Ask the material clarification questions and wait.
Keep the transcript unchanged. Do not build or invent human approval.
Step 3

Record the Human Check and prepare the handoff

Use only after a person reviews the exact spec version and chooses Proceed. If the decision is Revise or Pause, do not build.

Proceed with spec [actual version] for this local fictional prototype,
including these reviewed assumptions: [list or refer to their IDs].
Record this decision and its version faithfully. Create outputs/build_handoff.md
for Codex. Include the original task, paths to the source, clarifications and
approved spec, the exact approved version and decision, acceptance check IDs,
exclusions and unresolved questions. Do not invent reviewer details or approval.
Do not build. The next step will run separately in Codex.
Step 4

Codex builds the approved prototype

Switch to Codex in the same extracted project folder. The files are the handoff.

Read .agents/skills/spec-to-prototype/SKILL.md and use that skill.
Read outputs/build_handoff.md, the complete referenced spec, source and
clarifications. Verify that the human Proceed applies to this version and
scope. If absent, stale or contradictory, stop and identify the missing decision.
Build outputs/prototype/index.html as one offline HTML file with sample data.
Use a polished, readable layout on laptop and phone; keep the approved scope.
Run available browser checks and save actions, observations, failures and
untested checks in outputs/build_notes.md. Do not infer passes from code alone.
Create outputs/verification_handoff.md with the exact spec version, prototype
path, how to open it, evidence paths and limitations. Do not approve your own
outcome, contact anyone or publish. Return the local preview.
Step 5

Fresh Claude Code reviews the outcome

First give a person the original task without button-by-button instructions. Record only what actually happened.

Read .claude/skills/validate-outcome/SKILL.md and use that skill.
Read outputs/verification_handoff.md, the original transcript and task,
outputs/spec.md and outputs/clarifications.md. Inspect the prototype and
recorded evidence yourself; do not adopt the builder's verdict.
If browser tools are available, exercise the checks and record new observations.
If not, identify which supplied evidence supports each finding and leave
unsupported checks Not tested. Label agent-operated checks separately from
intended-user observations. Here are the actual volunteer observations:
[paste observations, or state that no volunteer has tested it].
Save outputs/validation.md with Pass, Fail or Not tested for every check,
the main mismatch or evidence gap, and the smallest next action.
Do not edit the prototype, change requirements, invent approval or publish.
Next step

Build this workflow for your own task

In AI Sandbox Lab, you bring a real work task and take it from context and specification to a verified tool. Group 7 starts on November 17 and is taught in Russian.