Agent frameworks are multiplying like end-of-quarter OKRs.
Every week a new way to wire tools, memory, graphs, swarms, and “autonomous” loops. Some are thoughtful. Some are README-driven development with a logo.
My reaction to the new-framework headline cycle: curiosity yes, migration panic no.
Why the Cambrian explosion is happening
The core model capabilities jumped. Tool calling got better. Developers smelled a platform. Investors smelled a platform. Everyone ships a thin orchestration layer and calls it a category.
This is normal. It is also noisy.
Under the noise, most teams need a boring loop: plan a little, act a little, check results, stop when done, keep logs.
You can build that without marrying a framework religion on day one.
What to evaluate before you care
- Escape hatch. Can I eject to plain code without archaeology?
- Observability. Can You see traces, tool calls, costs, and failures clearly?
- Control. Hard stops, budgets, human gates.
- State model. Explicit state beats mystical memory.
- Community vs lock-in. Is the value the idea or the prison?
If a framework cannot answer those, it is a demo wrapper.
Opinion: your problem is probably not the framework
Most agent failures You see are:
- unclear task definitions
- missing evals
- secrets in the wrong place
- tools that are too broad
- no ownership for outcomes
A new graph library does not heal that. It can decorate that.
Fix the job design first. Then choose orchestration sugar if it still helps.
A sane adoption path
- Start with one serious workflow, not twelve toys
- Implement a minimal loop in plain code
- Add a framework only when it removes real pain (retries, tracing, tool plumbing)
- Keep business logic outside the framework temple
- Re-run evals on every framework upgrade
Upgrades are where frameworks tax you. Budget for the tax.
Checklist for the next shiny release
- Does it solve a pain we logged before the tweet?
- Can we spike in a day without rewriting the company?
- Are logs first-class?
- Is there a kill switch?
- Would we still want the patterns if the library died?
The quiet winners
The quiet winners will look boring in screenshots: clear tools, tight permissions, good evals, simple orchestration, strong operational habits.
The loud winners will look magical for a month. Some will become real. Many will become migration lore.
Closing
Agent frameworks are multiplying fast. That means the ecosystem is excited, not that you are behind.
Stay calm. Steal ideas. Delay commitment until a framework earns it on your workflow with your metrics.
The best agent stack is the one your team can debug at 11pm without reading a novel of abstractions. Keep that as the north star while the logos fly by.
Steal patterns, not prisons
Sequence diagrams, retry policies, tool allowlists, and budget caps are portable ideas.
You can learn them from a framework without rewriting your stack into that framework. Ideas travel. Lock-in stays home unless you invite it in.
Closing
Let the frameworks multiply. Let your standards stay picky. Ship the workflow you can explain, debug, and afford.
Team skill portfolio
Invest in skills that outlive libraries: writing briefs, designing tools, building evals, reading traces. Those skills make any framework optional. Optional is power.
One workflow deep
It is better to run one workflow deeply with traces and evals than tour eight frameworks shallowly. Depth creates taste. Taste makes the next framework choice faster because you finally know what pain you are buying your way out of.
Closing note
Frameworks are multiplying. Your attention should not. Pick pain, pick proofs, pick portable patterns.
One more pass
When a new framework drops, write three bullets: what pain it claims to remove, what lock-in it introduces, and what proof you would need before a spike. If you cannot fill the bullets, you are not evaluating. You are scrolling.
Bottom line
Stay picky. Ship anyway.
