Sharkly

Manage AI coding agents like teammates, not tabs. One board for tasks, parallel runs, blockers, and review across Claude Code, Codex, and more.

[Sharkly @SharklyAI]

Your team already runs AI coding agents. What it doesn't have is a way to manage them like teammates.

Sharkly is that layer: assign a bug straight to Claude Code or Codex, run every task in its own git worktree on your own machines, and review results on a real board. The agents execute. People set direction and accept the work.

We're building it in public here.

[Sharkly @SharklyAI]

@aresotik Persistent sessions are the easy win. harder: when you come back, how do you know what to review? the agent ran for 6 hours, touched 40 files, and wrote a commit message. that handoff is still mostly manual

[Sharkly @SharklyAI]

@arpit_bhayani Review load is where it breaks in practice. 5 sub-agents, 5 open PRs waiting on the orchestrator. Planning overhead disappears; code review time stays linear. You traded one bottleneck for another

[Sharkly @SharklyAI]

Claude Opus 5.5 is live. Perplexity WANDR benchmark: 0.610 score, $4.13 per task, 67.6% cheaper than Fable 5.1 at similar output. For teams running coding agents, per-task cost is the number that governs your compute budget, not per-MTok.

[Perplexity @perplexity_ai]

Claude Opus 5.5 is now available as the Standard effort level in Perplexity Computer.

On our WANDR benchmark, it scored 0.610 at $4.13 per task, slightly outperforming Claude Fable 5.1 while costing 67.6% less per task.

[Sharkly @SharklyAI]

@beamnxw The 'coordinate the work' plugins are per-session. once you have 3 agents running in parallel the problem moves up a layer: task ownership, which run is stuck, two agents targeting the same file. plugins don't see that. that layer has to live above the IDE.

[Sharkly @SharklyAI]

@atmoio Spec length is layer one. the wall most people hit next is coordination: three agents in flight, nobody watching the stuck run, two of them editing the same file. The spec problem is solvable with discipline. The second one needs structure

[Sharkly @SharklyAI]

https://t.co/HN6Rex9GkD

[Sharkly @SharklyAI]

OpenAI: 3.1 agent-workdays per human workday inside its research org, and over half of successful 4-8 hour tasks needed a human to step in.

Do the math. Four agents running means two or three of them stop and wait for you per task, at times you didn't pick.

The queue that grows at 3:1 is results waiting for a person. Most tooling doesn't show it. Where does yours land?

[Sharkly @SharklyAI]

https://t.co/TFUkvSKjPO

[Sharkly @SharklyAI]

Three kinds of credentials in an agent setup, and they shouldn't mix:

1. your own API token (shk_)
2. a single-use install token (sit_)
3. the scoped credentials each agent runs with

Agents never log in as you. If one leaks, you know exactly what it could reach.

[Sharkly @SharklyAI]

This is the right comparison for picking your first agent. once you're running both in parallel, the problem shifts: who reviews each run, who owns a file two agents touched, and which one is sitting blocked on feedback. different question entirely.

[Paul Hudson @twostraws]

Having now spent enough time with both Fable and Astra, here's how I've found Codex and Claude compare for app development, along with some tips for success 👇 https://t.co/6E16eBUL1n

[Sharkly @SharklyAI]

Worktrees solve the file-conflict problem. the coordination problem is different: who reviews an agent's PR, what happens to the stuck run nobody is watching, what order do parallel worktrees merge in. these are org problems, not isolation problems.

[Rhys @RhysSullivan]

it's the best time in the world for new tools!

my agents use rift instead of git worktrees, no idea how it works i just told codex i wanted to use it

if anything the incumbents are at a disadvantage, look at how much the models struggled with nextjs changes

[Sharkly @SharklyAI]

https://t.co/1C0tUd85e2

[Sharkly @SharklyAI]

Your agents are going to interrupt you. The only question is where.

An inbox that carries just 'waiting for you' and 'failed' beats a wall of logs and a Slack channel that never sleeps.

[Sharkly @SharklyAI]

@BradGroux @Muse @AIatMeta @Meta @MetaforDevs 12/15 failures being non-code is the pattern every multi-agent run hits. The fix is isolation: one workspace per agent, explicit handoffs between them. Shared state (auth contexts, working dirs, daemons) is where running seven-in-parallel falls apart.

[Sharkly @SharklyAI]

@AliFarahat Interesting

[Sharkly @SharklyAI]

Projects ships parallel threads inside one runtime. when you mix claude code, codex, and gemini cli, the coordination problem just moved up: who owns the work when three runtimes finish at once and nobody set up a handoff?

[ClaudeDevs @ClaudeDevs]

Today we're rolling out Projects in Claude Code on desktop and web.

A project is one conversation with Claude. It splits the work into threads itself, runs them as parallel cloud sessions, passes context between them, and keeps going when you leave.

In beta for select users.

[Sharkly @SharklyAI]

https://t.co/xCRHuGvx7B

[Sharkly @SharklyAI]

Agentic coding is the agent taking a task, not a prompt.

The difference shows up in what you have to hand over: a scope, a repo, a place to run, and a definition of done. Miss one and you're back to prompting.

[Sharkly @SharklyAI]

@addyosmani Parallel threads is the easy part. the hard part hits when three threads want to touch the same module. Who reviews? what order do those PRs merge? you need a board, not just an orchestrator.

[Sharkly @SharklyAI]

https://t.co/uo0sImMxLK

[Sharkly @SharklyAI]

There's a special grief in coming back to a laptop that slept through your agent's run.

Sharkly Desktop has a prevent-sleep toggle with durations from 5 minutes to Always, plus a separate keep-running-with-lid-closed option on macOS (it asks you to plug in first). Small toggle. Large amount of un-wasted compute.