Managing agent context with Tessl

[KRRT7]

Hey Guido,

Following up on our call — you mentioned struggling with agents.md: too much in it wastes tokens, too little and the AI makes bad decisions (or cheats against benchmarks).

I've been working with a tool called Tessl that handles this. It manages agent context — what instructions get loaded, when, and how much — so you're not hand-tuning a single markdown file and hoping for the best.

Would you be open to trying it on TypeAgent? I can set it up and show you what it looks like in practice.

[gvanrossum]

How do I use Tessl? The website you pointed at looks like it's just marketing. (Too loud for my conservative taste.)

…

--Guido

On Tue, Apr 28, 2026 at 5:21 PM Kevin Turcios ***@***.***> wrote: Hey Guido, Following up on our call — you mentioned struggling with agents.md: too much in it wastes tokens, too little and the AI makes bad decisions (or cheats against benchmarks). I've been working with a tool called Tessl <https://tessl.io> that handles this. It manages agent context — what instructions get loaded, when, and how much — so you're not hand-tuning a single markdown file and hoping for the best. Would you be open to trying it on TypeAgent? I can set it up and show you what it looks like in practice. — Reply to this email directly, view it on GitHub <#257>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWCWMVSJ2QUSVT266YMHHT4YFDINAVCNFSM6AAAAACYKCIK42VHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZZHE3TGMBUGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[KRRT7]

Tessl is a CLI that manages context for AI coding agents (Claude Code, Cursor, Copilot, Gemini, etc.). It works through "tiles" — versioned bundles of instructions, rules, and docs that get loaded into your agent's context automatically.

The key difference from hand-editing AGENTS.md: tiles are versioned, scoped, and agent-agnostic. It also automatically scaffolds the config for each tool — Claude Code, Gemini, VS Code Copilot, Cursor, etc. — and centrally manages all of it in one place, so a change auto-applies across every agent.

Docs: https://docs.tessl.io
Installation: https://docs.tessl.io/introduction-to-tessl/installation

You'll have to create an account to use it — not sure if that'll put you off, but it's a powerful tool.

Happy to set it up on TypeAgent if you want to see it in action.

[gvanrossum]

I don't feel I have a problem that requires such a complex tool. Sorry.

…

--Guido

On Tue, Apr 28, 2026 at 7:19 PM Kevin Turcios ***@***.***> wrote: Tessl is a CLI that manages context for AI coding agents (Claude Code, Cursor, Copilot, Gemini, etc.). It works through "tiles" — versioned bundles of instructions, rules, and docs that get loaded into your agent's context automatically. The key difference from hand-editing AGENTS.md: tiles are versioned, scoped, and agent-agnostic. It also automatically scaffolds the config for each tool — Claude Code, Gemini, VS Code Copilot, Cursor, etc. — and centrally manages all of it in one place, so a change auto-applies across every agent. Docs: https://docs.tessl.io Installation: https://docs.tessl.io/introduction-to-tessl/installation You'll have to create an account to use it — not sure if that'll put you off, but it's a powerful tool. Happy to set it up on TypeAgent if you want to see it in action. — Reply to this email directly, view it on GitHub <#257 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAWCWMSO5ROGNP6AY5J3QCT4YFREZAVCNFSM6AAAAACYKCIK42VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMNZVGA2DOMI> . You are receiving this because you commented.Message ID: ***@***.***>

[KRRT7]

Fair enough — but looking at your current AGENTS.md, I think the problem is already there, just not obvious yet.

The file mixes 6 different concerns into one flat list: behavioral tone, git restrictions, build commands, uv rules, code generation style, and meta-instructions. Everything loads on every interaction, whether relevant or not.

A few concrete things:

• "NEVER COMMIT CODE" is a text instruction the AI can choose to ignore. Both Claude Code and VS Code Copilot support hooks that would actually block the command from running.
• "Don't use ! on the command line" is a bash workaround — irrelevant to agents that don't run shell commands, but it loads there too.
• The code generation section (~60% of the file) loads when you ask "why is this test failing."
• "When I state a general rule, add it to AGENTS.md" burns tokens on every single request.

PR #256 adds more (worktree rules, debugging discipline). The file will keep growing, and every rule loads every time.

This doesn't have to be Tessl. If you're open to it, I can start small with your current VS Code setup — add a PreToolUse hook that actually enforces the "never commit" rule, and split the code generation conventions into a scoped file that only loads when relevant. That way you can see the benefit without adopting anything new.

[KRRT7]

To make this more concrete, I put together a draft PR: #258

It does three things:

1. Splits AGENTS.md from 106 → 32 lines — code generation conventions move to .github/code-conventions.md, only loaded when writing code instead of every interaction
2. Replaces "NEVER COMMIT CODE" text with a PreToolUse hook — actually blocks the command instead of hoping the AI listens. Works with both Claude Code and VS Code Copilot.
3. 39 pytest tests verifying the hook blocks mutations and allows read-only commands

No new tools required — just files already in the repo.