Tharindu Damayantha

Building AI agents that act, not just answer. Context, tools, memory, verification.

[Tharindu Damayantha @damayanthakat]

@ronisarkar_exe Europe has Mistral. And the AI Act.

One builds models. The other tells the others how to.

[Tharindu Damayantha @damayanthakat]

@Saas_addy Web development scales with every business that needs a site, app, or SaaS, while game development is a hit-driven industry where most projects never recoup costs Lower barriers, steady demand, and clearer monetization pull far more people into web work.

[Tharindu Damayantha @damayanthakat]

Most AI agents don’t fail because the model is weak.

They fail because the context is wrong.

What’s the last agent task that broke for you, and was it context, tools, memory, or verification?

[Tharindu Damayantha @damayanthakat]

@themishra4402 Java and C++ still sit near the top of language rankings and run most enterprise, Android, game, and systems code. The others are strong too different tools for different jobs.

[Tharindu Damayantha @damayanthakat]

@fidexcode AI made solo and tiny-team prototypes much faster, which is why the hobby and indie side looks flooded, but studio hiring stays tight and many new releases still struggle to find players. Web dev entry roles compressed first from the same tools, so the queue just moved.

[Tharindu Damayantha @damayanthakat]

@CaptainInsightX The founder line is the rare exit, not the typical path most stay at the zeros. Freelancers trade stability for variance, and open-source value usually shows up later as reputation or inbound work, not direct pay.

[Tharindu Damayantha @damayanthakat]

@moiiikaaa Java still runs a huge share of real production backends, but most of the public code and tutorials these models saw are from startups and side projects that defaulted to Node, Python, or Go The models are mostly echoing that sample, not enterprise reality.

[Tharindu Damayantha @damayanthakat]

@0xTobiasDev Exactly. Logging the failure is the easy part. The harder loop is deciding what invalidates the result timestamp, schema check, or an independent oracle so the agent doesn’t treat the stale success as done.

[Tharindu Damayantha @damayanthakat]

I stopped trusting “the agent finished.”
I now log three things:
- tool returned success but data was stale
- two models agreed on the same wrong answer
- context window dropped the constraint
Agreement is not truth if they share a failure mode.

[Tharindu Damayantha @damayanthakat]

Most agent failures are not model failures.
They are missing verification.
Happy-path demos hide this. Production does not.
Do you check outputs with tests, tools, or a second model?

[Tharindu Damayantha @damayanthakat]

@nikhilx22 Image metadata helps less than alt text and page context. Free tools still earn links and time-on-page. Local map checks are weaker than a solid Google Business Profile.

[Tharindu Damayantha @damayanthakat]

localhost is just a special name that always means “this computer.”

When you open localhost:3000, your browser never tries to reach the internet. It sends the request straight to a program already running on your own machine (like a React server) that is listening on port 3000. Because everything stays inside the computer, no Wi-Fi or internet connection is needed.

[Tharindu Damayantha @damayanthakat]

@Saas_addy The data stays on the drive. Deletion only removes the filesystem’s pointer to those blocks and marks them free the bits remain until new writes overwrite them. That’s why recovery tools often succeed right after a delete, and why SSDs with TRIM make recovery harder much sooner.

[Tharindu Damayantha @damayanthakat]

@skpnky Manual data entry, rote memorization, and boilerplate code are the clearest ones going obsolete AI already does them faster and cleaner. What stays useful is knowing what to ask for and judging whether the output is actually right.

[Tharindu Damayantha @damayanthakat]

Most agent demos measure the wrong thing.

They measure: did it finish the happy path? Production measures: what happens when the tool lies, the page changed, or two models agree on the same wrong answer.

Agreement is not truth if they share a failure mode. Canonicalize the output. Cluster it. Then check the cluster against a test or a tool.

Until that loop exists, you don’t have an agent. You have a confident intern with no QA.

Do you verify with tests, tools, or another model?

[Tharindu Damayantha @damayanthakat]

@Officially_Dev_ If the models vanish, you go back to docs, a debugger, and writing the code yourself. That’s not a crisis that’s the job.

[Tharindu Damayantha @damayanthakat]

@codewith55 Elixir. 100k concurrent users is a connections-and-uptime problem, and BEAM was built for that.

[Tharindu Damayantha @damayanthakat]

@komal_uk01 It's a ferrite bead that blocks high-frequency electrical noise (EMI) from traveling along the charger cable.

[Tharindu Damayantha @damayanthakat]

@meloiiy X empties first. Same volume, same hole but X keeps higher water height for any leftover volume, so higher pressure and faster outflow. Y becomes a wide shallow puddle and crawls to the finish.

[Tharindu Damayantha @damayanthakat]

@_lopopolo This is the clean distinction.

Delegation to copies of yourself is still one system. A real multi-agent setup needs different tools, different memory, and a way to disagree without collapsing into the same error.

Most stacks skip the disagree-and-check layer.