DEV Community

Cover image for Anthropic Just Validated Agent Teams, Why Specs Matter More Than Prompts
Anton Abyzov
Anton Abyzov

Posted on

Anthropic Just Validated Agent Teams, Why Specs Matter More Than Prompts

#opensource

Anthropic Just Validated Agent Teams, Why Specs Matter More Than Prompts

Today Anthropic showed the slide that matters most for the next phase of agentic software: Agent Teams Revisited.

Not because it is flashy, but because it makes the shift explicit:

  • subagents
  • Claude Code communicating with them
  • coordination as a first class capability

That is the real story.

Most people still think in terms of one super prompt and one super assistant.
I think the future is much closer to an organization:

  • one lead agent
  • multiple specialists
  • a shared spec
  • verification before execution

Prompts give output, specs give alignment

A prompt is good for producing a result.
A spec is better for aligning a system.

Once you want multiple agents to work together, alignment matters more than cleverness.
Without structure, multi agent workflows become multi chaos.
With structure, they become leverage.

That is why I am bullish on spec first orchestration.

The workflow I want is simple:

spec -> /sw:team-lead -> specialists -> verification
Enter fullscreen mode Exit fullscreen mode

The lead agent should not improvise from vague intent.
It should have enough structure to:

  1. understand the target outcome
  2. break work into tasks
  3. delegate to specialists
  4. review what came back
  5. return a clean execution path

Why Anthropic's slide matters

When a frontier lab shows subagents and Claude Code communicating with each other, it validates a broader direction:

  • better models will matter
  • but orchestration will matter just as much
  • the next moat is not only intelligence, it is coordination

A lot of the value will move into the workflow layer around the model.
That includes how work is specified, delegated, verified, and merged.

What I am building toward

This is the direction I have been pushing with SpecWeave and the Verified Skill layer.

Both are FREE and OPEN SOURCE:

My view is simple:

  • prompts are not enough
  • teams need specs
  • agent teams need a lead
  • execution needs verification

Anthropic showed the destination.
Now the real race is building the best execution layer around it.

If you are building in this space, I would pay very close attention to that shift.

Top comments (0)