Evot.AI

Execute. Observe. Evolve. The open source runtime for self-improving AI agents.

[Evot.AI @Evot_AI]

Evot is open source here:
https://t.co/ZXrtDnOtza
/harden is built in now — use it on plans, designs, or current git changes before shipping.

[Evot.AI @Evot_AI]

Inspired by @cjzafir’s “loophole loop”, Evot now has `/harden`.
It stress-tests a plan or your current git changes, finds concrete loopholes, proposes fixes, and repeats until the strategy holds.
It’s a simple way to get more out of frontier models like GPT-5.5 and Opus 4.7: don’t just agree — break the strategy, patch it, then re-check it.
Try /harden before you implement, commit, or ship.

[Evot.AI @Evot_AI]

https://t.co/ZXrtDnOtza

[Evot.AI @Evot_AI]

Evot's long context compaction uses zero LLM calls ​—​ pure heuristics + static analysis:

Spill: results >12KB → disk, 4KB preview in context

L0: clear expired results · shrink oversized → tree-sitter structural outlines for code (Rust/Python/JS/TS/Go), head+tail truncation for text, per-tool policies

L1: collapse old turns into [Summary] tool list ​—​ "key text" via pattern extraction

L2: drop middle messages, keep head+tail, tail-first retention

No summarization model. No extra API cost. Images never stripped. 2+ consecutive errors with >50% budget → auto-compact + retry.

[Evot.AI @Evot_AI]

https://t.co/XAL6fADtm0

[Evot.AI @Evot_AI]

https://t.co/gy5N0WWJbQ

[Evot.AI @Evot_AI]

Try: https://t.co/ZXrtDnOtza

[Evot.AI @Evot_AI]

Evot now supports full-text search across ALL sessions!
Type any keyword to instantly find and resume your past conversations.
Just try: resume + search

[Evot.AI @Evot_AI]

https://t.co/Hye488ksp6

[Bohu @BohuTANG]

⚠️ Official @OpenAI GPT-5.4 (Pro Plan) is secretly injecting 𝐚𝐝𝐬 / 𝐮𝐧𝐡𝐞𝐚𝐥𝐭𝐡𝐲 𝐜𝐨𝐧𝐭𝐞𝐧𝐭 into tool results.
The Codex agent never sees it — the second the context looks “not invalid,” it just tells the LLM to retry and buries the garbage.
But I caught it red-handed with @Evot_AI . 😎

[Evot.AI @Evot_AI]

Evot’s First Security Barrier just activated.

Every skill is scanned at install — read-only analysis only.
Zero execution. Zero injection risk.
Risks are exposed instantly. No surprises. No trust required.
Security isn’t optional — it’s the first gate.

[Evot.AI @Evot_AI]

BendClaw has evolved into 𝐄𝐯𝐨𝐭.
Same engine. Same mission — self-evolving AI agents, fully observable, zero wasted tokens.

No LLM summaries. No external memory services. Context stays sharp through deterministic compaction — we don't ask the model to compress itself.

Everything ships under https://t.co/0rDfNvfpPM now.
Built with 🦀 Rust. 20k+ lines. Moving fast.

[Evot.AI @Evot_AI]

We just shipped BendClaw — the first agent built from the ground up for complete observability.

A self-evolving agent for long-running, high-complexity work. Continuous feedback loop:
observe execution → refine context → evolve toward the task.

Every token is now fully observable and quantifiable. In coding, read_file on massive codebases once ate 75%+ tokens and polluted context. Across any scenario, you see exact token distributions and compression patterns.

What makes it different:
• 𝙁𝙪𝙡𝙡 𝙤𝙗𝙨𝙚𝙧𝙫𝙖𝙗𝙞𝙡𝙞𝙩𝙮 — every token distribution visible in real time
• 𝙎𝙚𝙡𝙛-𝙚𝙫𝙤𝙡𝙫𝙞𝙣𝙜 — continuously improves its own context

Fewer tokens. Higher quality. By design.

Try it now → https://t.co/NuGULVbzQV

[Evot.AI @Evot_AI]

BendClaw — https://t.co/aS4zx5zuU3’s self-evolution harness runtime.
Forces agents to Execute → Observe → Evolve every run for true autonomous evolution.
https://t.co/gKudt8Rcj9

[Evot.AI @Evot_AI]

Self-Evolution Harness: 𝙀𝙭𝙚𝙘𝙪𝙩𝙚 → 𝙊𝙗𝙨𝙚𝙧𝙫𝙚 �� 𝙀𝙫𝙤𝙡𝙫𝙚

No prompt engineering. Just pure self-improvement.

1. Execute session (LLM + tools)
2. Snapshot BEFORE + runtime events
3. Observe replay
4. Evolve: extract knowledge
5. Inject learnings → restart

LOOP until improved.
Every run makes every agent smarter.
Rust-native • Shared memory • Real self-evolution.

[Evot.AI @Evot_AI]

https://t.co/79WK5PpY7X just shipped AI Tasks ⚡️
You don’t need to do anything — just chat.
The AI creates the full Task UI for you.
All you do is approve.

[Evot.AI @Evot_AI]

Trace and visualize every operation -- who did what, which tools fired, what came back. All queryable:

[Evot.AI @Evot_AI]

We just shipped 𝙀𝙫𝙤𝙩.𝘼𝙄, and open-sourced the runtime behind it.

It's a 𝙙𝙞𝙨𝙩𝙧𝙞𝙗𝙪𝙩𝙚𝙙 𝘼𝙄 𝙩𝙚𝙖𝙢 𝙥𝙡𝙖𝙩𝙛𝙤𝙧𝙢 built for teams and enterprises. Multiple agents running in parallel, sharing the same context and memory, getting smarter together over time.

Here's the problem we kept running into. You give one agent a complex task, it gets slow, it loses context halfway, and it starts from scratch every time. The obvious fix is more agents. But then they're all islands -- Agent A figured out your deployment process last week, and Agent B has no idea.

So we built a 𝙨𝙝𝙖𝙧𝙚𝙙 𝙙𝙖𝙩𝙖 𝙡𝙖𝙮𝙚𝙧 underneath. Every agent runs on its own node with isolated compute, but they all read and write to the same memory, the same knowledge base, the same history. When one agent learns something, it's there for everyone on the next run. A brand new agent on day one already has everything the team has ever figured out.

What makes it different:
- 𝙎𝙝𝙖𝙧𝙚𝙙 𝙗𝙧𝙖𝙞𝙣 -- all agents connected to one data layer. Context, memory, knowledge, history, not siloed.
- 𝗗𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗲𝗱 𝗯𝘆 𝗱𝗲𝗳𝗮𝘂𝗹𝘁 -- agents dispatch subtasks across nodes. One overloaded? It hands off to a peer.
- 𝙇𝙤𝙘𝙖𝙡 + 𝘾𝙡𝙤𝙪𝙙 -- local nodes touch your files and dev environment, cloud nodes handle heavy compute. One cluster, no boundaries.
- 𝙎𝙚𝙡𝙛-𝙚𝙫𝙤𝙡𝙫𝙞𝙣𝙜 -- every completed run deposits knowledge back into the shared pool. No prompt tuning, no manual handoff.
- 𝙁𝙪𝙡𝙡 𝙩𝙧𝙖𝙘𝙚 𝙖𝙣𝙙 𝙖𝙪𝙙𝙞𝙩 -- every operation recorded end to end. Who did what, which tools fired, what came back. All queryable.
- 𝙋𝙧𝙤𝙙𝙪𝙘𝙩𝙞𝙤𝙣-𝙧𝙚𝙖𝙙𝙮 -- secrets vault-isolated, approval gates on sensitive actions, per-agent token budgets with proactive alerts.
- 100+ 𝙞𝙣𝙩𝙚𝙜𝙧𝙖𝙩𝙞𝙤𝙣𝙨 -- Slack, Lark, GitHub, Linear, and counting. Internal tools? Write a Skill, point it at a script. Done.

Humans set direction, review what matters, approve what's critical. 𝙀𝙫𝙚𝙧𝙮𝙩𝙝𝙞𝙣𝙜 𝙞𝙣 𝙗𝙚𝙩𝙬𝙚𝙚𝙣 𝙞𝙨 𝙩𝙝𝙚 𝘼𝙄 𝙩𝙚𝙖𝙢'𝙨 𝙟𝙤𝙗.

The engine underneath is 𝘽𝙚𝙣𝙙𝘾𝙡𝙖𝙬 -- distributed AgentOS, written in Rust, open source as of today.

Come try it: https://t.co/6NECxlFzOs
Star the repo: https://t.co/M1Ta2UPwhH